How VED Saarthi works

Multi-turn chat with GPT-5.5 and Claude Sonnet 4.6. Switch models mid-conversation using the model dropdown in the composer. Free users get 50 messages/day on GPT-5.5; paid users get unlimited chat on both models.

• Drag-and-drop chats between Projects in the left sidebar.
• Chats stream in real time (Server-Sent Events).
• Long paste / typing auto-expands the composer without overflowing.
• Enter sends, Shift+Enter for new line.

Claude Opus 4.8 is Anthropic's top-tier reasoning model, sitting a rung above Sonnet in depth and nuance. Pick it from the Intelligence pill (composer, bottom-right) when your prompt is worth the extra thinking budget — it's slower and pricier per token, but the answers show it.

When to reach for Opus vs Sonnet:

• Long-form drafting — reports, proposals, essays, legal-style writing.
• Complex code review — architecture critiques, security audits, tricky refactors.
• Multi-step planning — breaking a fuzzy goal into ordered steps with trade-offs.
• Nuanced analysis — comparing options, weighing pros/cons, spotting edge cases.
• For everyday chat, translation, quick Q&A → stay on Sonnet 4.6 (faster + cheaper).

Availability: Opus is a paid-tier premium model — visible with a Pro pill in the model picker. Free-tier users see the option but selecting it triggers the upgrade nudge. Pro and Power subscribers can pick Opus on any chat.

Why Opus 4.8 (not Opus 5)? Anthropic released Claude Opus 5 on Jul 24, 2026. VED Saarthi runs on the Emergent Universal LLM key which routes `claude-opus-4-8` today; the moment Opus 5 is onboarded upstream, we'll flip the model ID and your existing Opus chats will seamlessly upgrade — no action needed on your end.

Tip: Combine Opus with the High Intelligence setting (same pill) for maximum reasoning effort — great for gnarly problems, but keep an eye on your monthly token budget in Settings → Credits & Usage.

Gemini is Google's LLM family, available in three sizes on VED Saarthi. Pick any of them from the Intelligence pill (composer, bottom-right) — all three are labelled with a Pro pill because the Gemini tier is a paid feature.

Which Gemini should I use?

• Gemini 3.1 Pro — flagship. Best-in-class 1M-token context window: paste huge docs, transcripts, entire codebase excerpts. Smart router auto-picks this when your prompt crosses 2,000 characters.
• Gemini 3.5 Flash — latest fast tier. Balanced speed + quality for everyday chat when you want a different “voice” than Claude / GPT.
• Gemini 3 Flash — previous fast tier, kept available as a fallback if 3.5 Flash is rate-limited upstream.

Availability: Gemini is a paid-tier feature — Free users see the option but selecting it triggers the upgrade nudge. Pro and Power subscribers can pick any Gemini size on any chat, and the smart router will transparently pick Gemini 3.1 Pro for very long prompts.

Tip: Gemini shines when your prompt is huge — a 2,000+ char transcript, a whole PDF pasted in, or a long codebase excerpt. For short questions Claude Sonnet / GPT-5.5 are still faster and cheaper. The composer shows a one-tap nudge when you cross the 2,000-character mark.

Vision: All three Gemini sizes accept image uploads through the same paperclip attachment flow — the [+] menu → “Add photos & files” button. Great for “analyse this screenshot / diagram / document” use cases when you want a Google-flavoured second opinion alongside GPT-5.5's vision output.

Compare model outputs lets you send one prompt to multiple models in parallel (GPT-5.5, Claude Sonnet, Gemini, Opus — any mix of 2-4), then pick the reply you like best. Only the winner is saved to your chat — the losing replies are silently discarded.

Why you'll love it:

• No more “which model should I pick?” — let all of them answer, then choose.
• Faster than sending the same prompt N times — requests run in parallel, so total wait ≈ slowest model (not sum of all).
• Zero chat clutter — only your picked reply is stored. No 3 assistant bubbles to scroll past later.
• Latency chip on each card — tells you exactly how long each model took, so you can weigh speed vs quality.
• Failure-tolerant — if one model errors out (rate limit, timeout), the others still show. You're never blocked.

How to use:

1. Type your prompt in the composer (don't hit Send).
2. Click the [+] button on the left of the composer.
3. Pick “Compare models” from the menu.
4. Check 2-4 models in the picker, hit Run compare.
5. Read all replies side-by-side, click Use this reply on your favourite.

Tip: Great for creative writing (pick the model with the best “voice”), tricky code (see whose fix is cleanest), and any question where you don't know upfront which model will nail it. Premium models (Opus, Gemini) still need a paid plan to compare against.

When the composer input crosses 2,000 characters (about ½ a page of text), a small teal chip appears above the composer suggesting Gemini 3.1 Pro. Gemini has a massive 1M-token context window, which is far more headroom than Claude's 200K or GPT-5.5's 128K — so long pastes stay intact instead of getting silently truncated.

Why the nudge exists:

• Better answers on long docs — the model actually reads your whole paste instead of losing the tail.
• One-tap switch — hit the Switch to Gemini button, chip disappears, done. No model-picker digging.
• Non-intrusive — X out and it stays gone for the rest of your session. Drop the input below 1,800 chars and it resets so it's ready for your next long paste.
• Free-plan aware — if you're on the free tier the button becomes See plans instead, so you always know what's available.

Related: above 6,000 characters you also get a Preview summary button in the same chip — see the “Prompt preview” help section below.

Prompt preview shows you what a long paste actually says before you hit Send. When your composer input crosses 6,000 characters, a small Preview summary button appears inside the long-prompt nudge chip. Click it, wait ~5-15s, and you'll get:

• TL;DR — a 120-180 word summary of the whole paste in a teal card at the top.
• Section list — up to 8 chunks with a 40-80 word summary each. Split intelligently by Markdown headings (##, #) or paragraph groups.
• Original preview — click any section to expand and see the raw text that summary is based on.
• Char stats — footer tells you what percentage of the original the TL;DR is (usually 5-15%).

Two ways to send:

• Keep original — dismiss the modal, your full paste stays in the composer, send normally.
• Use TL;DR instead — replaces your composer input with the 120-180 word summary. Great when you realise most of the paste was noise. Confirms first so you don't lose the original by mistake.

Why you'll love it:

• No surprises — see exactly what the model will read before you spend the tokens.
• Catch repetition / padding early — the TL;DR calls out sections that are just repeated boilerplate.
• Save tokens on huge pastes — if the TL;DR captures your intent, the “Use TL;DR instead” button lets you send 10% of the chars for the same answer.
• Uses the cheap summarizer (Claude Sonnet 4.6) — preview cost is a tiny fraction of what your real model call would use.

Tip: Best for pasted transcripts, PDF text dumps, long codebase excerpts, or anywhere you copy-pasted more than a screenful. If your prompt is under 6,000 chars, just send — the model can handle it without a preview step.

New chat — start a fresh conversation. Each chat has its own memory and shows up in the Conversations list below.

Generate image (white button) — opens the image generator. Describe what you want (e.g. “a serene mountain landscape at sunrise”) and pick GPT-Image-1 or Gemini Nano-Banana.

Brand poster style (Style picker) — inside the image generator, pick the VED Saarthi Poster style to produce a square 1:1 marketing infographic in our brand palette — V logo, hero headline, pricing pill, 5 feature cards, Start Free CTA, URL footer. Just describe your offer in the prompt; the art direction is baked in. Also available: Infographic (clean explainer) and Illustration (flat vector).

Projects (group related chats) — click + New next to Projects in the sidebar to create a folder (e.g. Goa Trip, Q3 Planning). Use the + New chat in project shortcut inside each project to start a chat that already lives there. The project ⋯ menu has Instructions — a custom system prompt that is prepended to every chat in that project (think of it as persistent context for an ongoing workstream).

Drag-and-drop a chat into / out of a project — grab any chat in the sidebar and drop it onto a project row to file it there. Drop it onto the Conversations header (or the dashed “Drag a chat here to move it out of its project” zone) to take it back out. The chat itself is never modified — only its project assignment.

Multi-select — move 10+ chats at once — hold Shift and click two chats to select a contiguous range, or hold Ctrl / ⌘ to toggle individual chats in/out. A “N selected” toolbar appears — drag any of the selected rows onto a project (or back to Conversations) to move them all in one go. A plain click anywhere clears the selection and opens that chat.

Mermaid diagrams inside chat — paste any ```mermaid code block in your prompt — or ask the assistant for one — and it renders inline as a proper diagram. Click the diagram to expand it in a lightbox; use the Export PNG / Export SVG buttons to save it to disk.

⌘+K — Insert diagram from a template — press ⌘+K (or Ctrl+K) anywhere in chat to open the Insert Diagram modal. Pick a starter (flowchart, sequence, ER, mindmap, gantt) and edit live — the preview updates as you type. Insert at the composer or save as the chat's pinned diagram.

Pin a diagram as your “Living Spec” — use the Pin button on any rendered diagram to lock it to the top of the chat. While it's pinned, the diagram source is passed to the LLM as living context on every reply — so the AI always answers consistent with the current spec. Edit it any time from the pinned header; unpin to remove the context.

The Settings page lives at /app/settings and is organised into four panels via a left sidebar (mirrors ChatGPT for familiarity). Deep-link any panel via a hash — for example /app/settings#voice opens the Voice panel directly.

• General — Response management: default reply length, knowledge level, answer style, and whether to include examples / code. Per-chat overrides on the composer pill always win.
• Voice — pick the TTS voice used when the assistant reads a reply aloud. The picker moved here from the composer so all audio preferences live in one place; Free plans get 1 voice, Pro 6, Power 9.
• Billing — current plan card with renewal date, an annual-billing upsell, and an Upgrade shortcut to /pricing. Invoice history is coming soon; email support@vediautodevops.com for any past invoice in the meantime.
• Data controls — jump into the Archived chats view, open Shared-links guidance, or nuke everything with Delete all chats (double-confirm: JS prompt + typed “DELETE”). Full-account export is coming; per-chat Markdown export lives in the chat's ⋯ menu today.

Microphone — talk instead of type — tap the mic to record a voice query. We transcribe with Whisper and send the text to the LLM. Good for long, structured questions you'd rather speak than type.

Vision — drop an image to ask about it — paste, drag, or attach an image into the composer. The model sees the image and can describe it, extract text (OCR), explain charts, answer follow-ups about it, etc.

Chat kebab menu (⋯) — top-right of the active chat. Opens actions for the current thread: View files in chat, Move to project (hover-submenu), Pin chat, Archive, and Delete (with confirm).

Disclaimer placement — the “VED Saarthi can make mistakes” line now floats above the composer and hides automatically when you scroll up to read history, reappearing the moment you land back at the bottom.

Voice picker moved — the TTS voice selector used to sit next to the mic; it now lives in Settings → Voice.

“How to download” chip — a subtle “? How to download” link sits beside the AI disclaimer under the composer. Tap it to open a visual guide with screenshots of every download / copy button in the app — no need to type “download” to see the tutorial.

Talk to VED Saarthi with your voice. Tap the mic in the composer to record; the transcript is sent through the same chat + safety pipeline. On paid plans you can also enable hands-free continuous mode.

Pick your TTS voice — open Settings → Voice to change the voice used for read-aloud. The change takes effect on the very next reply and syncs across tabs.

• Pro: 100 voice messages / month, 6 TTS voices
• Power: 500 voice messages / month, all 9 voices incl. British storyteller

Drop a photo, screenshot, or document into the composer (the [+] button, drag-and-drop, or paste) and ask questions about it. Works for handwritten notes, product photos, code screenshots, receipts, whiteboards, and more. Pro plans get 300 images/month, Power gets unlimited.

File uploads: besides images, you can attach PDF, Word (docx), Excel (xlsx), CSV, TXT, Markdown, JSON and 20+ code file types (max 10 MB). VED extracts the content and answers questions about it — summaries, data analysis, error-spotting, rewriting, anything.

Multi-file upload: pick or drag several files at once — the first one lands in the composer chip so you can add a caption, and any extras get queued (shown as a +N queued pill next to the chip). After your first message goes through, the queued files auto-send one by one with a “Please analyse this file” prompt so you never have to click Upload five times in a row.

Ask directly in the search bar — say things like “Please create an image of an Azure DevOps logo” or “Draw a superhero cat wearing sunglasses”. VED auto-detects the image intent and renders a high-quality picture using GPT-Image-1 (with auto-picked portrait / landscape / square sizing based on your prompt). No modal, no options — the image just appears in the same chat bubble, ChatGPT-style.

Edit the last image with a follow-up — once VED has generated an image in the current chat, your next message can just say things like “make the background transparent”, “add the tagline ‘Deploy Faster’ below the logo”, or “change the color to green”. VED detects the edit intent, feeds the previous PNG back into GPT-Image-1's edit endpoint, and streams the updated version in place — no re-uploading, no modal, no options.

One-tap “Regenerate” chip — every AI-produced image now shows a small Regenerate pill below the picture. One click re-runs the same prompt (or the last edit) with a slight seed variation, appending a new bubble so you can compare takes side-by-side and pick your favourite. Great when a logo is 90% right and you want a couple of quick variations before committing.

“Compare N takes” lightbox — after two or more regenerations on the same prompt, a Compare N takes chip appears next to Regenerate on the latest bubble. Tap it to open a full-width dark lightbox showing every take side-by-side. Click any thumb, hit Pick this take, and the others quietly collapse into a 🗂 N alternates · show pill above your favourite — one tap on the pill un-hides them again. Keeps your chat tidy without losing history.

Prefer the modal? Click the white “Generate image” composer button to open the classic picker with style, size, and Nano Banana / GPT-Image-1 selection. Both flows count against the same monthly quota (Lite/Pro 300–400/mo, Power unlimited). Free tier does NOT include image generation.

Where: open the Create Image Spark and upload / generate a reference image. In the prompt textarea, type / — a menu opens with matching commands. Use ↑↓ to nav, Enter or Tab to pick. Commands that need extra info (e.g. /replacebg) prefill the prompt so you can type the scene; the rest run instantly.

Three tiers — colour-coded in the menu:

• Nano Banana (Tier 1) — 39 style / mood / design commands: /cartoon, /anime, /sketch, /watercolor, /oilpaint, /comic, /pixelart, /cinematic, /dramatic, /sunset, /night, /rain, /snow, /fog, /smoke, /fire, /lightning, /neon, /glow, /3deffect, /glass, /shadow, /watereffect, /reflection, /replacebg, /portrait, /headshot, /studio, /productshot, /mockup, /sticker, /poster, /thumbnail, /banner, /logo, /avatar, /infographic, /diagram, /architecture.
• fal.ai specialised (Tier 2 · BETA) — for the ops where dedicated models beat general LLMs: /removebg (transparent PNG via BiRefNet-class model), /upscale (Real-ESRGAN 2×/4×), and /restore (photo-restoration with scratch removal + colour fix).
• Client-side (Tier 3 · free) — /resize 1080x1080 resizes on Canvas instantly, no API call, no quota hit.

Quota: slash commands count against the same daily Create Image cap as regular generations, except /resize which is free. Free tier: 5 / day. Paid tiers: 20-100 / day depending on plan. Power tier: unlimited.

Reference image required: most commands need an image to work on. Upload one via the composer or pick from your gallery. Client-only /resize works on whatever image is currently in the composer preview.

Two ways to launch. Open the full editor from the Sparks Gallery card or type /ppt <topic> in any chat composer — the topic auto-fills. Aliases work too: /deck, /slides, /presentation.

Two generation flows.

• Sketch outline — quick 6-10 slide skeleton (just titles + one-line intent). Use this when you want to review the flow before committing tokens to the full deck.
• Generate full deck — one-shot: every slide gets a title, 3-5 impact bullets, and 2-3 sentences of speaker notes.

Per-slide AI rewrite. Not happy with one slide? Hit the Rewrite pill on that slide — the model rewrites just the title, bullets and notes for that slide, keeping the rest of your deck untouched.

Preview before export. Once saved, hit the Preview button in the header. A full-screen modal renders every slide at the exact 13.33 × 7.5 in / 16:9 geometry your .pptx will use — WYSIWYG guarantee. Use ←/→ or the numbered strip to navigate; export .pptx or PDF straight from inside the preview.

Auto-polish — Make it Pro. The fuchsia Auto-polish button (enabled once the deck is saved) sends your slides to Claude Sonnet 4.6. It rebalances every bullet (≤5 lines, ≤14 words each), recommends a template that fits your deck's vibe, and proposes an image prompt per slide so any missing hero images can be one-tap generated from the diff view. Before / After preview + Apply or Discard.

Speaker Mode (Present). The indigo Present button opens a full-screen presenter view: current slide, next slide preview, speaker notes, per-slide + total timers. Set a Slide Budget (min/slide) and the SLIDE timer colour-codes green → amber → red as you overshoot 75% / 100% of the target — persists across sessions. Keyboard: ←/→ or Space to navigate, T resets timers, Home/End to jump, Esc to exit.

6 templates + Custom. Pick from the left rail:

• Business — navy + sky, clean corporate look
• Startup — violet + amber diagonal, pitch-deck energy
• Academic — Georgia serif, centered, conference-ready
• Creative — pink + orange block, agency vibe
• Minimal — Apple-style whitespace, no accents
• Storytelling — burnt-orange serif with drop-cap numbers
• Custom (your own) — hit Create custom template under the picker. Clone any built-in, tweak 5 colours + 2 fonts + layout, live-preview both cover and content, and save under your account. Personal templates show with a fuchsia Custom badge and can be deleted from the same row.

Per-slide images — 4 layouts. Attach an image to any slide (upload, pick from your Library, or AI-generate with a preset prompt). Then choose one of four layouts from the pill under the thumb:

• Compact — text left, small image right (default).
• Half — bullets left 50%, big image right 50%. Magazine style.
• Wide — bullets top, wide image below.
• Full — image fills the slide, bullets hidden. Best for hero slides.

Reference image for AI generation. In the image picker's AI generate tab, upload a reference from your device or pick one from your Library — the generator will edit / extend that picture using your prompt (image-to-image). Works with both GPT Image 1 and Gemini Nano Banana.

Export in either format. Hit .pptx for a native PowerPoint file (with speaker notes in Presenter View) or PDF for a printable / share-friendly version. Both use the template you picked and stay in sync with the on-screen preview. One-click download — save + export happens in the same click.

Public share link. Click Share in the header (enabled once the deck is saved) to publish at /p/ppt/<slug> — a clean read-only preview anyone can open, no login. Perfect for pitching before the .pptx is finalised. Click again to un-publish.

My Decks + drag-to-reorder. Every save shows up in the "My decks" list on the left rail with search, multi-select, bulk delete, and click-to-open. Slides can be reordered by dragging the grip handle or using the up/down arrows.

Tips for great decks.

• Add a subtitle + author for a polished title slide
• Keep bullets under 14 words — the exports look best when text doesn't wrap
• Use Rewrite with a hint like "add data" or "shorten"
• Run Auto-polish as the last step — one tap tightens every slide
• Full-layout hero slides read best with no bullets — Auto-polish enforces this automatically
• Set the Speaker Mode budget to 2 min/slide for a 20-min pitch

What is a “Spark”? A Spark is a small, focused AI helper built for one job — like “write my resume” or “make a poster”. You don't need to know how to prompt an AI. Just open the Spark, fill in a couple of boxes, and get a finished result.

Where to find them. Click Sparks Gallery in the sidebar, or open any of them directly:

• Resume Builder — write a resume that passes ATS bots
• ATS Checker — upload any resume, get a score + fixes
• Cover Letter — one letter tailored to a job description
• LinkedIn Bio — crisp 3-line profile bio
• Interview Prep — 15 questions + practice with your voice
• PPT Spark — AI presentations with speaker notes
• Ad Poster — Instagram / LinkedIn ad in seconds
• Create Image — text → picture with reference support

Pin your favourites. Every Spark has a pin icon. Pin up to 10 to your left sidebar so they're always one click away. Pins sync across your devices — pin on laptop, they show up on mobile too.

Slash catalog (power users). Type a slash in any chat to trigger a Spark without leaving the conversation — e.g. /resume, /ppt marketing plan, /ad instagram sale, /image sunset over Mumbai. Full catalog with search + tooltips is at the bottom of the Sparks Gallery — 44+ commands grouped by category.

Free tier limits at a glance.

• 5 resumes / day · 5 PPT decks / day · 5 ATS checks / day
• 2 AI images / day · 2 cover letters / day
• 1 interview prep, 1 roast / day
• Pro removes every cap. See Pricing.

Who this is for. Anyone applying for a job — freshers, students preparing for placements, working professionals switching roles, freelancers building a portfolio. No design skills needed.

Get started in 3 steps.

1. Open Resume Builder and answer a short quiz (or paste text in Hinglish — the AI understands).
2. Pick a template. Free gets 5 clean templates, Pro unlocks 7 premium ones.
3. Edit any bullet, add roles, hit Export PDF.

What's inside — the mini-tools on your resume.

• AI bullet rewrite — click the sparkle icon next to any bullet, add a hint like “quantify results” and get 3 rewrites. Free: 3 rewrites/day. Pro: unlimited.
• ATS score — the top-right panel scores your resume for common ATS filters (keywords, section headings, spacing). Fix flagged issues in one click.
• Tailor for a Job Description — paste any JD (up to 8,000 chars) and get a match score + missing keywords + actionable suggestions. Then click Quick Radar for a live before/after keyword tracker.
• Cover Letter Companion — one-click generate a 250-350 word cover letter grounded in your resume. 4 tones: professional / warm / bold / concise.
• Interview Prep — get 15 questions bucketed as behavioural / technical / resume deep-dive, each with a hint on what the interviewer is actually testing. Then hit Practice on any question, speak your answer using your phone / laptop mic, and get an AI score + strengths + a sample strong answer.
• Roast Mode — feeling brave? An affectionate ~200-word roast of your resume with every jab paired with a fix. Great for a reality check before submitting.

Free vs Pro.

• Free: 5 saved resumes · 5 templates · 3 AI rewrites/day · watermark on PDF
• Pro: Unlimited resumes · all 12 templates · unlimited rewrites · no watermark · unlimited JD Match + Interview Prep

Share it. Every resume has a public URL you can send to a recruiter — read-only, no login needed. Toggle it off any time.

Why this exists. Most companies use ATS (Applicant Tracking System) software to filter resumes before a human ever reads them. A resume that looks great to you might get rejected because of a missing keyword, weird formatting, or a date that doesn't add up. This tool checks all of that in ~20 seconds.

How to use it — 3 steps.

1. Open ATS Checker.
2. Drag & drop a PDF or DOCX file (up to 5 MB).
3. Wait ~20 seconds. Read your score, fix the red flags.

What you get.

• Score /100 — colour-coded: green (80+), amber (60-79), red (<60). Section-wise: Contact, Summary, Experience, Skills, Education.
• Grammar checks — typos, awkward phrasing, weak verbs. One-click “Apply all fixes”.
• Authenticity flags — future start dates, impossibly short/long tenures, title/company mismatches.
• Rewrite for a JD — every flagged role has a purple pill that opens Cover Letter Spark pre-filled with that role & company, so you can tailor for a specific job.

Free vs Pro. Free: 5 uploads/day, 3 auto-fix rewrites/day. Pro: unlimited on both.

Tip: Export a copy of your Resume Builder resume as PDF and run it through here — you'll see exactly which keywords a recruiter's ATS filter will miss.

Who this is for. Anyone applying to a job that asks for a cover letter — or wants to stand out with one even when it's optional.

3-step flow.

1. Open Cover Letter Spark. It pulls your most recent resume automatically.
2. Type the role + company. Paste the JD (optional but recommended).
3. Pick a tone — professional, warm, bold or concise — and hit Generate.

JD Keyword Radar (built-in). Right below the letter, you get:

• A traffic-light match % — green ≥ 70, amber ≥ 40, red < 40.
• Missing keywords as red pills — with a tip to weave 2-3 verbatim into your bullets.
• Matched keywords as green pills — already covered by your resume.

Recent runs. Every generation is saved to a “Recent runs” strip — click any chip to reload that letter + its keyword radar. Great for A/B-testing tone or role.

Free vs Pro. Free: 2 letters/day. Pro: unlimited.

What it does. Turns rough notes about what you do into a punchy 3-line profile bio you can paste directly into LinkedIn's About section (or Twitter, Instagram, portfolio — anywhere a bio is asked for).

3-step flow.

1. Open LinkedIn Bio.
2. Type your role, industry, and 2-3 achievements or things you care about.
3. Hit Generate. Copy, paste, done.

What makes it different. The AI is instructed to avoid cliches (“passionate about”, “results-driven”, “synergy”) and write like a real person. If the first version feels generic, hit Regenerate for a different voice.

Two modes in one Spark.

• Text → Image. Just describe what you want — e.g. “a chai tapri at 6 AM with steam curling off the kettle, warm morning light”.
• Image → Image. Upload a reference photo (up to 5 MB) or pick one from your Library. The AI edits / extends it using your prompt. Great for “change the sky to sunset” or “make this a Pixar-style version”.

Intent chips — one-tap presets. The row above the prompt box has 6 chips (Hero image, Diagram, Team photo, Chart mockup, Background, Concept metaphor). Tap one to inject a battle-tested starter prompt + auto-select the matching style.

Diagram style — Kubernetes / Docker-quality architecture. Pick Diagram as your style (or just type “create docker image”, “kubernetes architecture”, or any of ~20 tech keywords) and the backend wraps your brief in a scaffold — header bar, labelled rounded-rectangle boxes with icons, solid arrows for data-flow / dashed for control-flow, a colour-coded legend, and a key-benefits row. Aspect auto-locks to 16:9 because diagrams look terrible in portrait.

Diagram Templates gallery — 6 tiles. When you're in Diagram mode, an indigo collapsible surface exposes: Kubernetes, AWS VPC, Microservices, CI / CD, ML Pipeline, Data Flow. Each tile is a rich ~500-char prompt iteratively refined against Nano Banana — tap once, tweak the details, hit Generate.

Guide me — turn a vague brief into a rich one. Type a rough diagram idea (“how do I show my ci/cd pipeline?”) and hit the indigo Guide me pill next to Suggest richer prompt. Powered by Claude Haiku (cheap, fast), it returns a suggested richer prompt + 4-6 key components to include + 2-3 relevant in-app Sparks (/ppt, /course, /task) with a one-line "why". Tap Use this prompt to swap the textarea contents.

Editable Prompt Description panel. After generation, an indigo collapsible panel below the image shows the exact rich description sent to Nano Banana (with a Diagram mode badge when it fires). Tweak the copy, hit Regenerate with description, and iterate on the AI's own richer text — much stronger signal than the 1-liner you typed.

Image action bar. Under every generated image, a ChatGPT-style toolbar:

• Comment — jot a private note (kept locally per image).
• Resize — client-side canvas with 5 presets (1080×1080 IG, 1920×1080 16:9, 1200×628 Social, 800×600 Slide, 512×512 Thumbnail), width/height inputs, lock-aspect toggle, and Copy / Download.
• Remove BG + Erase — visible but greyed out — coming soon once the fal.ai plug-in is wired in.

Chat integration — smart nudges + /diagram slash.

• Type anything diagram-y in chat and an inline banner slides in: “VED Saarthi can create even better images from this.” Tap Get rich description to fetch a scaffolded prompt + components, or Open in Create Image → to jump into the Spark pre-filled.
• The nudge also fires for logo, poster, thumbnail, avatar, hero, mockup, and infographic intents — copy adapts per intent.
• /diagram k8s, /diagram aws, /diagram microservices, /diagram cicd, /diagram ml, or /diagram data-flow in the chat composer opens Create Image Spark in a new tab with the picked template already filled in — zero LLM call, zero friction. Free-form tail works too: /diagram my ci pipeline for a nextjs app.
• Every image generated in chat gets a post-message Spark Upgrade CTA card below it — “Want an even better version? Open in Create Image Spark…” — deep-linked to the Spark with the prompt already filled in.

Gallery — S / M / L / List views + sort + multi-select. Every generation auto-saves to your personal Gallery (last 30). Top-right controls:

• S / M / L / ☰ toggle — 4 density modes. Small packs 8-per-row (scan 30+ images fast), Medium is the 4-per-row default, Large shows a 3-per-row hero layout, and ☰ List is a Windows-Explorer-style row layout with thumb + prompt + style + aspect + date + inline actions. Choice persists in localStorage.
• Sort dropdown (all views) + clickable column headers in list view (Style / Aspect / Date) that cycle asc → desc → default with ↑ / ↓ arrows.
• Hover actions per tile: Reuse (loads the prompt into the composer), Download PNG, Share (mints a public URL), Delete.
• Select mode — top-right Select toggle enters multi-select: every tile gets a checkbox, header shows 18 saved · 3 selected, and the toolbar swaps in Zip N (bulk ZIP download via JSZip, purely client-side) + Delete N (bulk delete).

Public share URLs + iframe embed. Tap the sky-blue Share button on any gallery image to mint a public link like /p/diagram/c7JcopiolF that renders the image + prompt + attribution — no login required. Perfect for Slack / LinkedIn / X. Every public page has a matching /embed/diagram/<slug> route with an iframe-friendly bare viewer for Notion / Confluence / Craft docs. Add ?bare=1 to strip attribution.

Prompt Remixer. Not sure how to describe what you want? Type a rough idea and hit the fuchsia Suggest richer prompt pill — the AI expands it into a rich 2-4 sentence prompt covering subject, lighting, palette, framing and mood. Great for photo / cinematic / illustration intents. Use Guide me instead for diagrams.

Discover on the Sparks home. Two dedicated entry points on the Sparks Gallery: the Diagram Studio card (sky-blue) that jumps straight to Diagram + 16:9 + templates gallery, and a Recently drawn strip showing your 4 latest architecture diagrams — tap any to re-open, tweak, regenerate.

Free vs Pro. Free: 2 images/day. Pro: 20/day. Gallery, Templates, Guide me, Share URLs, Bulk ZIP, and all view modes are free on every plan.

What it is. Instead of writing paragraphs, type / to open the autocomplete popover with 40+ built-in commands like /interview, /lesson, /roadmap, /quiz, /debug. Each one expands into a polished expert-crafted system prompt — you get more consistent, structured replies than free-form asking.

Positional parameters. Most commands accept arguments in a fixed order. Examples:

• /interview linux 25 medium — 25 medium-difficulty Linux Q&As
• /lesson javascript closures — a structured lesson on closures
• /quiz python 10 hard — 10 hard Python quiz questions
• /compare python rust for web APIs — side-by-side comparison
• /roadmap devops 6 months — 6-month DevOps learning plan
• /diagram k8s — opens Create Image Spark with the Kubernetes architecture template pre-filled (also /diagram aws, /diagram microservices, /diagram cicd, /diagram ml, /diagram data-flow, or any free-form tail). Zero LLM call — jumps straight to the picture.
• /ppt AI ethics for startups — opens PPT Spark with the topic pre-filled

Keyboard navigation. ↑/↓ to move, Enter or Tab to pick, Esc to close. Once you type a space after the command name the popup hides — you're now in params mode.

Free vs paid. Free-tier users get the 8 highest-utility commands: /ask · /explain · /summarize · /quiz · /code · /lesson · /image · /doc. Paid tiers unlock all 40+. Locked rows still appear in the popover with a 🔒 icon and an Upgrade pill that routes to /pricing.

Define your own. Head to Settings → Slash commands → New command. Give it a name (e.g. standup), a description, an ordered comma-separated param list (e.g. name, yesterday, today, blockers), and a template that references each param inside {name}-style placeholders.

Once saved, every future chat can invoke it like a native slash command: /standup Alice bug-fixes new-feature waiting-for-designs. The template expands, the LLM sees the polished prompt, and your chat bubble shows a purple Custom badge.

Tier gates. Creating custom commands unlocks on Pro (₹199/mo) — up to 25 personal templates. Publishing those templates to the community marketplace stays a Power perk (keeps distribution tightly moderated). Free / Lite users can browse and fork any public template into their private library — a forked template becomes usable the moment they upgrade.

Marketplace — publish, rate & fork. Each custom command has a Publish toggle (Power+) that flips it public and assigns a stable template_id. Anyone browses via Settings → Slash commands → Marketplace, rates with ❤️, or Forks a template into their private library instantly. Sort by Most loved, Most forked, or Recently published.

Limits. Up to 25 custom commands per user; names must be lowercase alphanumeric (2-24 chars); built-in command names always win the parse so you can't shadow /ask.

Multi-channel campaigns in one prompt. Say /ad instagram Fresh coffee brand launch, 20% off and VED renders a 1080×1080 IG-optimised poster with an LLM-crafted 3-6 word headline and platform-appropriate CTA. Skip the platform argument (e.g. /ad Fresh coffee launch) and VED renders the 4-platform hero pack: Instagram + Facebook + WhatsApp + Newspaper — all with the same headline for campaign consistency.

Supported platforms & sizes.

• instagram · 1080×1080 square feed (alias: ig, insta)
• instagram_story · 1080×1920 vertical (alias: story, reel)
• facebook · 1200×630 landscape (alias: fb)
• whatsapp · 1080×1920 vertical status (alias: wa)
• newspaper · 1200×1600 portrait print (alias: print, news)
• linkedin · 1200×627 corporate landscape (alias: li)
• twitter · 1600×900 16:9 (alias: x)

Quotas. Free tier: 4 ad images per rolling UTC day. Each generated poster counts as 1 — a 4-platform hero pack exhausts the daily quota in one call. Paid tiers get unlimited ad generation (still respects the account-wide image_gen daily cap).

Persistence. Ad campaigns are saved into your chat with a headline + platforms metadata block on the assistant message, so you can revisit and download any poster later. Each image shows a floating platform badge (e.g. “Instagram feed · 1:1”) in the top-left so you know which channel it's sized for.

What it is. One command builds a full picture-book explanation of any topic:

• An excited one-sentence hook.
• 3 illustrated steps (bright cartoon art, thick lines, clean labels).
• A surprising fun fact at the end.

Under the hood. One Claude Sonnet call builds the lesson JSON + 3 Gemini Nano Banana images are rendered in parallel via asyncio.gather, so the whole picture-book lands in a single chat reply.

Syntax

• /teachme <topic> [age:5-7|8-10|11-13] — original picture-book mode (3 illustrations + lesson)
• /teachme short|medium|long <topic> — teacher-notes mode (same output as /teachnotes)
• (attach image) /teachme — teacher-notes about the image (image mode)
• /teachme [age:20+] [domain:medical|legal|engineering|finance|science|design|teaching] <topic> — full workshop-tone notes for professionals

Length words: short ≈ 250 words (compact reference), medium ≈ 600 words (full classroom handout with flow + benefits), long ≈ 1500 words (deep dive with prose paragraphs). Only kicks in when the length flag is the first word OR an image is attached — otherwise plain /teachme <topic> keeps its illustrated picture-book behaviour.

Age bands: 5-7 kids · 8-10 preteens · 11-13 early teens · 14-17 teens · 20+ adults/professionals. Bare-age shortcuts also work — /teachme short 25 auto-maps to the 20+ band.

Domain presets (14-17 & 20+): layer [domain:medical], [domain:legal], [domain:engineering], [domain:finance], [domain:science], [domain:design], or [domain:teaching] to skew examples + terminology to that field. See /teachnotes help below for the full list with aliases.

The [age:X-Y] marker is optional — leave it off and the composer applies your most-used age from previous sessions (or the default 8-10 on first run).

Age bands

• 5-7 — sing-song, 8-word sentences, only toy / snack / pet analogies. Ideal for early readers.
• 8-10 — short sentences with sport / food / video-game analogies (default).
• 11-13 — one or two real technical terms per step, immediately explained in plain English. Middle-school comfortable.

Examples

• /teachme photosynthesis (auto-picks your recent age)
• /teachme why the sky is blue [age:5-7]
• /teachme fractions [age:8-10]
• /teachme how airplanes fly [age:11-13]
• /teachme बादल कैसे बनते हैं (Hindi topic — reply stays kid-friendly)

Age chips in the composer. The moment you type /teachme, three floating age chips (5-7 / 8-10 / 11-13) appear above the composer pill. Tap one to lock the age before hitting Send. Your most-used band is remembered per device and pre-highlighted with a · recent pill so a single Enter uses it automatically.

🎨 Try a different picture-book. Every lesson reply has a re-roll button underneath it — one tap runs /teachme again at the same age for fresh illustrations, or use the age chips right under the message to instantly re-teach at a different age.

What it is. A one-tap way to build teacher-ready notes on any topic, structured so you can scan them in 30 seconds before class:

• 📖 Overview — 2-3 sentence intro framing what and why.
• 🎯 TL;DR — one sentence you can read aloud.
• 🧒 Simple explanation / 🎓 Working definition — tuned to the age band + domain you pick.
• 📚 Deep dive — 2-5 numbered sub-sections with prose intros + bulleted takeaways.
• 🔄 Key flow / process — arrow chain (Step A → Step B → Step C) when the topic has a natural sequence.
• 🌍 Real-world examples — 2-4 concrete analogies with 1-3 line explanations each.
• ✨ Key benefits / takeaways — 4-8 poster-friendly bullets.
• ❓ Discussion questions — 3 open-ended prompts.
• ⚠️ Common student mistakes — framed as "students often think X, but actually Y".
• 💡 Teaching tip — one activity you can run in the next 10 minutes.
• 📝 Student-friendly summary — one bold sentence a student could repeat back.

Grammar

• /teachnotes help — friendly onboarding walkthrough with tap-able examples
• /teachnotes <topic> — default long (~1500 words, deep dive)
• /teachnotes short <topic> — ~250 word compact card (2 sub-sections, quick reference)
• /teachnotes medium <topic> — ~600 word full classroom handout (3-4 sub-sections + flow + benefits)
• /teachnotes long <topic> — ~1500 word deep dive with full prose paragraphs (default)
• /teachnotes [age:5-7|8-10|11-13|14-17|20+] <topic> — explicit age override (kids · preteens · early teens · teens · adults/professionals)
• /teachnotes short 25 — bare number is treated as the audience's age (25 → adults/professionals band)
• Also works as /teachme short|medium|long <topic> — same behaviour, familiar name

Age bands at a glance:

• 5-7 — Kids · sing-song analogies, tiny sentences
• 8-10 — Preteens · everyday analogies (default)
• 11-13 — Early teens · real terms, gently explained
• 14-17 — Teens · high-school pacing, technical but accessible
• 20+ — Adults / Professionals · workshop tone, precise terminology, no kid analogies

🎓 Domain presets (for `14-17` and `20+` only):

Add a [domain:X] marker on top of your age band to swap the tone, examples, terminology and "common mistakes" hints for a specific professional field. Kid bands ignore the marker (we never want to quote a Supreme Court ruling to a 6-year-old).

• [domain:medical] — clinical cases, drug/anatomy terms, SBAR handovers, red-flag misdiagnosis pitfalls · aliases: med healthcare clinical nursing
• [domain:legal] — landmark rulings with jurisdiction, IRAC breakdowns, statute citations, precedent splits · alias: law
• [domain:engineering] — worked calcs with units, design constraints, standards (ASME/RFC/ISO), failure modes · aliases: eng software dev tech
• [domain:finance] — real deals, ratio walk-throughs (EBITDA, P/E, IRR), base-rate + anchoring pitfalls · aliases: fin business mba biz
• [domain:science] — named experiments, replication considerations, p-hacking watch-outs · aliases: research stem
• [domain:design] — Fitts's / Hick's Law, Nielsen heuristics, accessibility red flags · aliases: ux ui product
• [domain:teaching] — ZPD, Bloom's, formative-assessment prompts, plenary + exit-ticket ideas · aliases: edu pedagogy instruction

Domain examples:

• /teachnotes [age:20+] [domain:legal] short mens rea
• /teachnotes [age:20+] [domain:medical] medium acute pancreatitis
• /teachnotes [age:20+] [domain:engineering] long PID controllers
• /teachnotes [age:20+] [domain:finance] short discounted cash flow
• /teachnotes [age:14-17] [domain:science] medium mRNA vaccines
• (attach image) /teachme [age:20+] [domain:engineering] short — analyse a system diagram as an engineering worked example

Length words at a glance:

• short — ~100 words · 30-second summary · warm-up card
• medium — ~300 words · summarised lesson · 5-min read
• long — ~700 words · descriptive full prep · homework handout

The [age:X-Y] marker is optional — leave it off and the command reuses your remembered /teachme age band. First time only, three age chips appear in the reply so you can pick one with a single tap.

📸 Image mode

Attach a photo of a diagram, textbook page, worksheet or whiteboard, then type /teachnotes (topic optional). The AI describes what's in the image first, then builds the seven-section teaching card around it — perfect for turning a screenshot into a lesson.

Examples

• /teachnotes photosynthesis
• /teachnotes short water cycle
• /teachnotes long why the sky is blue [age:5-7]
• /teachnotes fractions [age:8-10]
• (attach diagram) /teachnotes
• (attach diagram) /teachnotes short focus on the water cycle

💾 Age memory. The band you pick is remembered per device (same store as /teachme), so subsequent /teachnotes sends skip the picker and jump straight to the notes.

📄 Handout export. Ask "export as pdf" (or "as docx") right after a reply and the doc-generator wraps the notes into a printable handout you can share on WhatsApp, email or your LMS.

What it is. A safe, client-side control panel for every slash that remembers something between sends (like /teachme's age band), plus your own custom picks — key-value pairs any prompt can reference via {{key}}.

Grammar

• /prefs — list every pref (built-in + custom)
• /prefs <slash> — drill into one built-in
• /prefs reset <slash> — wipe a built-in's memory
• /prefs set <key> <value> — create/overwrite a custom pref
• /prefs delete <key> — remove a custom pref
• /prefs clear — wipe every custom pref
• /prefs share <key> — copy a one-liner to give someone
• /prefs import <key>=<value> — paste a shared one-liner
• /prefs packs — browse curated starter-pack bundles
• /prefs preview <slug> — dry-run diff (writes nothing)
• /prefs install <slug> — one-tap install a pack
• /prefs uninstall <slug> — reverse an install (customisations preserved)

Custom picks (mini-agent). Any prompt can embed a stored value with {{key}}:

/prefs set favourite_style cyberpunk
/image a moody portrait, {{favourite_style}} lighting

The {{favourite_style}} marker is substituted with cyberpunk right before your message flies to the LLM. Unknown keys are left literal so a typo is easy to spot in your own bubble.

Share a pref with a friend — safe by design:

/prefs share favourite_style
→ copy the one-liner: /prefs import favourite_style=cyberpunk
(friend pastes it into their composer + hits Send)

The receiver's /prefs import runs the SAME validation as /prefs set — key regex, reserved-prefix block, 50-key / 500-char caps, sensitive-key warning. Nothing auto-executes and nothing travels through any server. The friend still has to paste + Send.

Preset packs — one-tap curated bundles:

/prefs packs                     browse the catalog
/prefs install cyberpunk-artist  install a curated pack

Bundles like Studio Ghibli painter, Cyberpunk artist, Product shot studio, and Homeschool parent kit drop in 2-3 {{key}} values you can reference immediately. Packs are shipped inside the app bundle (no server call, no third-party feed) and every value still goes through setCustomPref — every guardrail applies.

Rules & guardrails (why it's safe):

• Keys: ^[a-z][a-z0-9_]* (max 31 chars). Reserved prefixes blocked so you can't shadow internal storage keys.
• Max 50 custom picks per device, 500 chars per value.
• Substitution is single-pass and capped at 20 replacements per message — recursive or bomb-style inputs won't loop.
• 100% client-side. Nothing is sent to a server, nothing is synced across devices. Wipe it any time with /prefs clear.
• The app warns you if a key looks like password / api_key / token / secret — never store real credentials, they'd sit in localStorage in plain text.

Reset link on chip rows. The composer's /teachme age chips have a small "reset" underline when a memory exists — one tap does the same thing as /prefs reset teachme without leaving the composer.

What it is. A zero-cost thinking tool that turns a scratchy idea into a production-ready image prompt. Type /imagedesc a cozy home-office desk with plants and morning light and the assistant returns a structured card, not a wall of Markdown.

What the card contains.

• Four field chips: Category (Hero / Diagram / Team / Chart / Background / Concept), Diagram preset, Style (Photo / Illustration / 3D / Diagram / Watercolor / Line-art / Cinematic), and Aspect (1:1, 16:9, 9:16, etc.).
• A highlighted Prompt to paste block with a Copy prompt button that writes to your clipboard.
• A primary Open in Create Image Spark button that deep-links with prompt, style and aspect pre-filled — you skip the manual paste + form.
• A one-line reasoning footer explaining why those picks fit your idea (great for learning good prompt hygiene).

Community caching. Popular briefs are cached for 7 days and shared across all users, so repeat inputs return in under 100 ms with zero LLM spend. Also feeds the public community FAQ so highly-searched briefs become discoverable evergreen pages.

Examples

• /imagedesc exploded view of a mechanical wristwatch, studio render
• /imagedesc team photo, 6 engineers on rooftop, golden hour
• /imagedesc data-flow diagram: user → api → queue → worker → db

Two ways to use it.

• Slash: /travel delhi to jaipur 2 days or free-form asks like /travel weekend getaway from bangalore under 5k or /travel field trip for class 6 in pune.
• Spark page: open /travel for the full-screen map view — search origin + destination, pan/zoom, and see route pins on OpenStreetMap tiles. Zero paid Maps API calls.

What you get in the reply.

• Distance + time (road & rail if applicable), best route.
• Fuel + toll estimate in ₹.
• 3 must-see monuments / attractions along or at destination.
• Where to eat — 2 local favourites + 1 tourist-friendly option, with signature dish.
• Reliable fuel stops with highway + km marks.
• Best time to leave to avoid traffic + best season.
• One-line safety / weather / scam heads-up specific to the location.

Also handles holiday ideas (3 budget levels), school field trips (age-appropriateness + pack list), and city exploration (1-day & 2-day plans grouped by area).

Two content-pack flows.

• TikTok Creator Pack — one button on Create Image Spark generates 5 different cinematic reveals of the same subject (smoke, drone, 3D billboard, unboxing, product-in-hand) in parallel. Perfect for a full week of posts from one prompt.
• Multi-Size Ad Batch — one prompt renders 4 ads at the correct dimensions for Instagram, Facebook, LinkedIn and Google Display, all with the same headline for campaign consistency. Runs the four generations concurrently via asyncio.gather.

My Packs. Every pack you generate lands in /my-packs with thumbnails, generation timestamp and one-tap re-download of any variant. Nothing gets lost when a chat scrolls off the top.

Under the hood. Both flows use the parallel batch endpoints (POST /api/images/tiktok-pack and POST /api/images/ad-batch) — the client streams a progress rail so you see 5/5 or 4/4 tick as each render lands. Free-tier quota still applies to each generation.

Shortcuts. Instead of /ad instagram …, use:

• /adInsta — 1080×1080 Instagram square feed
• /adStory — 1080×1920 IG / WA vertical story
• /adFB — 1200×630 Facebook landscape
• /adGoogle — Google Display 300×250 / 728×90 variants
• /adLinkedIn — 1200×627 corporate landscape
• /poster — 1200×1600 print-quality portrait poster

Ad preset picker (UI). On the Create Image Spark composer, tap the Ad presets chip to open a picker that fills the platform, aspect ratio and starter template in one tap — no typing the slash. Great for non-power users.

Ad Template Shelf. Directly below the composer, a horizontally-scrollable shelf shows Marketplace-published ad templates (headline formula + prompt + style) that you can fork into your prompt with one tap.

Four cinematic presets, each mapped to a proven camera + lighting recipe:

• /explode <subject> — exploded parts view floating on a clean grid, callout arrows to each labelled component.
• /anatomy <subject> — cross-section with numbered labels, technical-illustration style.
• /cutaway <subject> — half the shell removed to reveal the interior, studio lighting.
• /threesixty <subject> — clean 360° turnaround-ready hero shot on seamless white.

Preset picker (UI). Create Image Spark also shows a Diagram presets chip below the composer — tap any tile to prefill the prompt with the correct camera recipe plus a sensible subject placeholder.

Cinematic reveal slashes (also free): /smokeReveal, /3dbillboard, /rain, /cinematic. These edit the most recent image you posted in the chat, no fresh upload required.

What it is. /faq is a public index of the most-hit community briefs. When multiple users run the same /imagedesc or /help query, the cached answer is promoted to its own /faq/<slug> page with JSON-LD schema — Google, Bing and Perplexity pick it up as an authoritative Q&A source.

Why it matters for you.

• Faster: cached briefs return instantly — no waiting for the LLM.
• Cheaper: the community carries the cost of the first generation; every subsequent visitor is free.
• Shareable: send a friend a stable URL, not a chat screenshot.

Pages are auto-refreshed weekly and evicted from the index when hit-count drops. Owner-only analytics at /admin/guide-cache tune the TTL and eviction thresholds.

The 12-lesson in-app course is a bite-sized walkthrough of every feature most users never discover on their own. Each lesson is 2-3 sentences plus 3-5 highlights, so the full tour is under 6 minutes end-to-end.

When it opens. Automatically once for new signups (about 800ms after you land in the chat, so it never crashes into an empty screen). Any time after that, reopen it from Settings → General → Take the 2-minute tour tile.

Navigation. ← / → arrow keys or Back/Next buttons; click any dot to jump; Esc closes and remembers your position. Progress syncs across devices — start on desktop, resume on mobile, VED knows exactly where you left off.

“Try it now” CTAs. Four lessons (chat basics, slash commands, image gen, and images) end with aTry /lesson…-style button that seeds the composer with a real command and closes the tour so you can hit Enter and see the feature immediately.

Whisper never emits a / glyph, so VED normalises spoken slash-command invocations into their canonical form before moderation or persistence. Any of these phrasings work:

• “hey VED, run slash interview linux 25 medium”
• “hey VED slash lesson javascript closures”
• “run slash quiz python 10 hard”
• “slash roadmap devops 6 months”

Safety net. The normaliser only fires when the extracted command matches a known built-in. Casual sentences like “What is slash notation in Python?” or “run slash foobar with args” fall through untouched — VED handles them as regular chat.

Works in both the single-turn mic (🎤) and continuous Jarvis mode (🎧). Your chat bubble shows the canonical /interview linux 25 medium form with the same teal slash pill that typed commands render.

Every image, screenshot, voice recording and document you attach is streamed to your private cloud vault — never inlined into the chat record. Chats stay lightning-fast to load and old attachments remain downloadable indefinitely (until you delete them yourself).

Voice-note playback: voice replies (your own recordings plus the spoken assistant messages you save) render as an inline audio player right in the chat bubble. Play, scrub, pause, or right-click → “Save audio as” to download the raw file for offline use.

File Manager (Settings → Data controls): shows total MB used vs. your plan's quota, with a colour-coded progress bar that turns amber at 70% and red at 90%. Tap View all to browse every file you've ever uploaded — thumbnail for images, MB size, upload date — and delete anything you no longer need. The trash icon removes the file from storage AND unsticks it from its chat message, so nothing orphaned is left behind.

Orphan cleanup: if you attached a file to the composer but never actually sent the message, it's tagged as an orphan and can be cleared in one tap via the Delete N orphans pill. Frees space without touching any file that made it into a real chat.

Older-files sweep: the Sweep older files panel bulk-deletes every attachment whose parent message is past a cutoff you pick (30 / 60 / 90 / 180 / 365 days). The two-step flow shows a preview first — how many attachments, how many MB freed — so nothing disappears until you confirm. Message text stays intact; only the binary files vanish.

Client-side quota guard: before an upload starts, VED checks whether it would push you over your tier limit. If yes, you get a friendly toast telling you how much you've used and where to free space — no wasted bandwidth, no failed uploads at the tail end of a 25 MB screenshot.

What is the rail? As soon as a chat has 3+ of your questions and you've scrolled at all, a column of short horizontal dashes appears flush against the right edge of the conversation. Each dash is one of your past queries, top-to-bottom in the order you asked them. The dash that is currently in your viewport is rendered as a bold dark bar — the rest of the queries are light grey ticks. It's a TOC for your own conversation.

Hover, click, navigate — hover any dash to see the first ~120 characters of that query in a dark tooltip — perfect for finding “the one where I asked about Razorpay webhooks” in a 40-turn chat. Click it to smooth-scroll the chat to that exact message. Click the bold active bar to do nothing — you're already there.

Full-featured task manager with AI parsing. Visit Tasks to get started. Owners have full access without a paid subscription.

• AI parse: type a natural-language line like “Remind me to submit tax invoice by Friday 5pm” or “Every Monday 9am — team standup”. VED extracts title, due date, priority, and recurrence in one click.
• Recurrence: daily, weekdays, weekly, monthly — the scheduler fires the reminder and re-creates the next occurrence automatically.
• Reminders: when a task comes due, a pop-up window appears in the app (with Mark done / Snooze 10 min / Snooze 1 hr buttons). Turn on Push (bell icon on the Tasks page) to also get browser notifications — they fire even when the tab is closed, including on Android. Push notifications include one-tap Snooze buttons too. (Email reminders are currently disabled.)
• Pop-up not appearing? Troubleshoot:
  • The app must be open in a browser tab and you must be signed in — the pop-up can't appear on a closed tab (use Push for that).
  • Reminders are checked every ~45–60 seconds, so a pop-up can arrive up to a minute after the due time.
  • Reminders require a task with a due date & time and status pending — completed tasks never remind.
  • Reminders fire once per due time. If you dismissed it, edit the task's due time or hit Snooze to get a fresh one.
  • Task reminders are a Pro / Power feature — on the Free tier the scheduler skips them.
  • Quick self-test: create a task due 2 minutes from now, keep the tab open, and wait — the pop-up should appear within ~3 minutes. If it doesn't, refresh the page and check you're still signed in.
• Status at a glance: tasks with a due date show a live pill — red “Overdue by 2 hrs” or green “Upcoming · in 3 hrs” — plus the full date & time (e.g. 3 Jul, 09:30 am).
• Triage ordering: the Pending list auto-sorts with the most overdue task at the top, then soonest-due; tasks without a due date sit at the bottom.
• Priority + tags: low / medium / high labels; up to 8 tags per task.
• Edit anytime: the pencil icon on any task opens the editor — change title, due date & time, priority, tags or recurrence.
• Snooze — presets, custom, favourites: tap the clock icon on any overdue task to snooze it forward. The menu offers 15 min · 1 hour · 3 hours · Tomorrow presets, a Custom input for any 1-1440 minute value, and (once you've used it) a Pinned section at the top for one-tap access.
  • Pin a duration: type the value in the Custom field and tap the pin icon beside it. Up to 4 pins are saved.
  • Reorder pins: drag any pinned row by its ⋮⋮ handle to re-arrange the menu — the most-used pin sits at the top.
  • Unpin: hover a pinned row and click the × that appears on the right.
  • Cross-device sync: pinned durations follow you across browsers and devices — they're stored on your account, not just this browser.
  • The reminder pop-up shows the same Pinned pills so you can one-tap-snooze from a fresh notification too.
• Export: one-click CSV download of all your tasks (opens in Excel / Google Sheets).
• Caps: Free/Lite 5 active · Pro 200 · Power 1,000. Owners: 1,000.

Every account starts on Free. Upgrade anytime from the sidebar (Subscription → Manage plan) to lift the daily caps and unlock extra features. Quotas reset on the calendar window shown below.

Switching tiers is instant — your existing quota carries over and the new caps apply from the next message. If you cancel, you keep access until the end of the billing period and then auto-revert to Free — your chats, projects, and pinned specs are never deleted.

Every reply from GPT-5.5 or Claude Sonnet 4.6 burns a small amount of tokens — think of tokens as the meter behind every message. Your monthly plan comes with a big pool of tokens; a live Usage pill in the sidebar and a detailed Settings → Usage & Credits panel show exactly how much you've used and how much is left this month.

What is a token? — a token is roughly ~4 characters of English (or ~¾ of a word). A short question like “What's the capital of India?” costs about 10 tokens; the reply costs another ~30. A longer “write me a 1-page email” round-trip typically costs 800–2,000 tokens combined. Attaching a file, a long chat history, or asking for a verbose reply all bump the count.

Your monthly pool — every account gets a fresh budget on the 1st of each month (UTC). Unused tokens do NOT roll over.

*Assumes a typical 1,000-token round-trip (your prompt + reply). Long documents, verbose answers, or big chat histories will consume more.

Warnings before you run out — in Settings → Usage & Credits pick a warning threshold (50%, 80%, or 90%). We'll show a toast the moment you cross it, so there's no surprise mid-month.

Quiet this month — Snooze warnings chip — if the toast is being noisy (e.g. you already know you're heavy this month and don't need reminders), tap the small Snooze this month chip next to the threshold picker. Warnings are silenced for the rest of the current billing period and un-mute automatically on the 1st of next month. The chip flips to “Muted till Mon Day” so you always see when they'll resume — tap it again to un-mute early.

Ran out mid-month? Instant top-up. — if you hit your cap, a Top-up ₹99 for 500k tokens button appears in the Usage panel. It uses Razorpay (UPI / card / net-banking), lands instantly, and the extra tokens stack on top of your regular plan for the rest of the current month.

7 easy ways to save credits

• Keep prompts short. “Summarise this in 3 bullets” costs a fraction of “write me a detailed report on…”.
• Set reply length to Short. Open Settings → General → Response management and switch Default reply length to Short — every reply gets ~50% smaller, so your pool lasts twice as long.
• Start a fresh chat for a new topic. Long chats re-send the entire history with every message. A new chat = a clean, cheap context window.
• Turn off “always include examples / code” in General if you don't need them — cuts each reply by 200–500 tokens.
• Prefer text over voice + vision when you don't need them. Attaching a photo or PDF adds hundreds of tokens on top of your prompt.
• Don't hammer “Regenerate”. Each regenerate is a fresh billable turn. Read the reply first, then ask a targeted follow-up (usually cheaper than starting over).
• Use a Project system prompt instead of pasting the same context every message. Project instructions add tokens once per turn, not manually on every message.

How VED compares — we run on the same top-tier models (GPT-5.5, Claude Sonnet 4.6, GPT-Image-1, Whisper, Nano Banana) that the big-name apps use, but at a small fraction of the price for an India-based user. A quick comparison at today's street rates:

That's roughly ~8× cheaper than ChatGPT Plus for Pro, and ~3× cheaper than ChatGPT Plus for Power — while still giving you GPT-5.5 and Claude Sonnet 4.6 in the same account, plus image generation, voice, vision, tasks, Projects and RAG memory. All billing in ₹, GST inclusive, 7-day money-back, cancel any time.

Third-party pricing shown for reference at ₹83/US$; check each provider for current rates. GST and card FX may add ~18–20% on top of US-billed apps.

Press ? anywhere inside /app to open the same cheat sheet as a floating modal (works everywhere except when you're typing in an input or textarea).

• ? — open the shortcuts cheat sheet

  Opens a floating modal listing every shortcut on this page. Ignored while you're typing so it never eats the character in the composer. Close with Esc.

• Esc — close any open menu, modal, or popover

  Universal escape hatch — closes the ⋯ chat menu, Insert Diagram modal, Files-in-chat panel, Mermaid lightbox, and the shortcuts cheat sheet.

• Ctrl / ⌘ + Shift + M — minimise sidebar footer

  Hides the plan, subscription and devices cards at the bottom of the left sidebar so the conversations list gets the reclaimed space. Press again to bring them back. The choice sticks across reloads; small-screen laptops auto-start minimised on first visit (you'll see a one-time toast teaching this shortcut the moment that happens).

• ⌘+K / Ctrl+K — insert diagram

  Anywhere in chat, opens the Insert Diagram modal — pick a starter (flowchart, sequence, ER, mindmap, gantt) and edit live with a previewing pane on the right. Insert into the composer or save as the chat's pinned diagram.

• ⌘+I / Ctrl+I — switch intelligence (model)

  Quick-flips the model picker (GPT-5.5 ↔ Claude Sonnet 4.6) without taking your hands off the keyboard. Same binding works even when you're mid-prompt in the composer.

• ⌘ + Enter / Ctrl + Enter — generate (inside Image modal)

  While the Image generator modal is open, fires the Generate button without reaching for the mouse.

• Shift + click / Ctrl / ⌘ + click — multi-select chats

  In the conversations sidebar, Shift-click two chats to select a contiguous range, or Ctrl/⌘-click to toggle individual chats in and out. A floating “N selected” toolbar appears — drag any selected row onto a project (or back to Conversations) to move the entire batch in one go.

Every request runs through a classifier before it reaches the LLM. Categories cover CSAM, weapons, phishing, hate speech, harassment, self-harm instructions, malicious hacking, privacy violations, and more (15 total). See our Acceptable Use Policy.

• First violation: warning email + refuse the request.
• Second violation: final-warning email + ops notified.
• Third violation (in a 30-day window): account permanently disabled.

Medical / legal / financial “should I take X?” questions are refused with a nudge to a licensed professional — these do not count as strikes.

Refund policy — 7-day money-back — if you're unhappy with a paid plan within 7 days of purchase, email us and we'll process a full refund — no questions asked. After 7 days we don't offer prorated cash refunds, but you can cancel at any time and keep access until the end of your billing period; the plan will then auto-revert to Free without re-charging.

Renewals — auto-renew unless you cancel — subscriptions renew automatically at the end of each period (monthly or annual, whichever you chose). The renewal charges your original payment method. Cancel any time from Settings → Subscription → Manage plan — cancellation stops the next renewal but doesn't pull access away mid-period.

Lost receipt or unrecognised charge? — drop us a line with the email you signed up with and the last 4 digits of the card / UPI ID and we'll send you a copy of the invoice (or the refund details, if you don't recognise the charge).

For all billing/refund questions email info@vediautodevops.com. Full policy at /refund-policy.

Each message you send and every reply you get is embedded (via text-embedding-3-small) and stored on your account. When you start a new chat, we search your OWN past messages for the ones most relevant to what you just asked and inject them into the system prompt as “relevant context from your past chats”.

• Folder memory — when a chat lives inside a project folder (e.g. “intervue”), VED also pulls recent snippets from the OTHER chats in that same folder. Ask “do you remember what I wrote in this folder?”and it will. Works instantly, even without embeddings.
• Strict per-user isolation — retrieval filters on your user_id and only searches chats you own. Zero cross-user leakage.
• Current-chat excluded — the RAG lookup skips the chat you're currently in (that history is already in context the normal way).
• PII redaction before the LLM sees your message — emails, phone numbers, cards, PAN, Aadhaar are masked with deterministic tokens ([PII_EMAIL_1] etc.) and un-masked in the reply. Your PII never leaves your account.
• 👍 / 👎 feedback on every assistant reply feeds an internal quality signal — used to spot regressions and improve prompts.
• “Remembering from your past chats” chip — when a reply drew on your history or folder memory, a small brain chip appears above it, so you always know when VED is using what it remembers.
• Kill switches — USE_RAG=false orUSE_PII_REDACTION=false in the backend .envinstantly disables either layer without a redeploy.

Every night at ~07:30 IST we take an automated gzipped JSON dump of your data — chats, messages, projects, tasks, subscriptions, payments, and user records. Backups are kept for 30 days.

• Location: /app/data/backups/veds_YYYY-MM-DD.json.gz
• Owner can list + restore from the Admin Portal (coming soon) or via the API endpoints below.
• Restore is a merge, not a replace — rows created after the backup was taken are untouched, so a restore can only add data, never remove it.
• Manual trigger: POST /api/admin/owner/backups/run
• List backups: GET /api/admin/owner/backups
• Restore: POST /api/admin/owner/backups/restore with {filename: "veds_2026-07-01.json.gz"}

Why was I signed out? — whenever we ship a new version of VED Saarthi, every active session is reset so you load the latest UI cleanly — no cached buttons, no stale state. Just sign in again with the same account; your chats, projects and settings are all preserved.

Page looks stale or buttons don't respond — we auto-detect new releases and silently refresh your tab. If you ever see something that looks off after an update, a one-time reload is enough — no need for a hard refresh. Use whichever shortcut matches your OS:

• Windows / Linux: Ctrl + R or just F5
• macOS: ⌘ + R

Your account data lives on our servers, never on your device, so nothing is lost during a reload.