← Back to SansadSaar

SansadSaar — User guide

A tour of the eight corpora SansadSaar surfaces, how to set up AI, how the privacy model works, and what costs (if any) come with each option.

1. What is SansadSaar?

A single-file browser app that opens up eight corpora of India's parliamentary and executive record — searchable, filterable, and (optionally) AI-readable in the same UI. Each corpus is scraped on its own schedule by an open-source GitHub Action and mirrored as static JSON on Cloudflare Workers. The app you're using runs entirely in your browser — there is no SansadSaar server, no account, no analytics.

Three things make it different from the upstream ParliamentWatch (whose Python scraper foundation powers the oversight corpora):

• It runs as a single HTML file — no Python, no install, no terminal.
• AI summaries and Q&A can run on your device (Gemma / Ternary Bonsai over WebGPU) — no API key required.
• If you want a stronger model, you can plug in your own API key (Anthropic, OpenAI, Gemini, Groq, OpenRouter, Ollama, or any OpenAI-compatible endpoint).

2. The eight corpora

The corpus chip strip near the top groups the eight datasets into four thematic buckets. Click any chip to switch the list and filter row to that corpus.

Header counts (X reports · Y with text) re-bind to whichever chip is active. The chip pill on the right ("DRSC: updated 4 hours ago") shows how fresh the mirror data is for that corpus.

3. The interface at a glance

Main view — top header, corpus chip strip, filters, sortable report list.

The header runs across the top:

• Header stats reflect the active corpus. Click a different chip and the numbers rebind to that corpus.
• AI status pill tells you what AI is configured. Click it to jump to Settings.
  • AI off — nothing configured yet
  • Loading model… — local model downloading or warming up
  • Gemma 4 E2B ready — local model ready to use
  • BYOK: Anthropic — your API key configured
• Settings (⚙) — pick the AI mode, load a model, configure web search, manage per-corpus deep-search toggles.
• Help (?) — what you're reading is the long-form version of that modal.

Below the header is the corpus chip strip: eight chips grouped under four labels (oversight, legislation, proceedings, executive). The active chip is highlighted; everything below it — filter row, list, status badge — belongs to that corpus.

4. Browse, search, filter

The filter row is corpus-aware — each corpus emits the filters that make sense for it. The shapes you'll see:

The Search box does substring match against titles by default. Open any report's Full text tab and that text becomes searchable too. To pre-fetch every extracted text for a corpus and search across its full body content, toggle Enable full-text search across <corpus> in Settings (size estimate shown — uses local cache after first fetch).

Each row also carries a status badge on the right:

• metadata — only metadata is published; no extracted text yet.
• text — full text has been extracted by the mirror and is searchable.
• summary — you've generated an AI summary that's cached in your browser.

Gazettes additionally carry an EN+HI badge when both English and Hindi text exist in the same record (the archive.org OCR is bilingual).

5. Reading a report

Report dialog — Details tab. PDF links go straight to sansad.in.

Click any row to open a four-tab dialog:

• Details — committee, report number, dates, links to the English and Hindi PDFs on sansad.in, source attribution.
• Full text — the extracted text from the PDF (when available). Searchable, scrollable.
• AI summary — generate a plain-English briefing of the report. Cached in your browser after the first run.
• Ask — chat with the report. The model answers based on the report's text (and your cached summary, when one exists).

6. Setting up AI

Click the ⚙ Settings icon in the header (or the AI status pill). Pick one of two modes:

Settings → Local AI section.

Option A — Local AI (default, no key, free)

Runs an open-weight model entirely in your browser via Transformers.js on WebGPU. The first load downloads weights from Hugging Face; the browser caches them so subsequent visits are instant.

1. Open Settings. AI mode = Local AI.
2. Pick a model from the dropdown.
3. Click Load. First run shows a progress bar — your machine is downloading weights from Hugging Face.
4. When the status pill says "<model> ready", you can use AI summary and Ask in any report dialog.

Option B — BYOK (Bring Your Own Key)

Settings → BYOK section. Provider list mirrors upstream ParliamentWatch.

Send the request directly from your browser to the provider you choose. Your key stays in this browser's localStorage — never sent to any other server.

1. Open Settings. AI mode = BYOK.
2. Pick a provider. Get a key from the provider's website (links in the table at §12 Costs & limits).
3. Paste the key into the API key field. Optionally override the default model.
4. Click Save. The pill turns to "BYOK: <Provider>".

7. Generating an AI summary

AI summary tab — a freshly-generated 4-section briefing.

1. Make sure AI is configured (Local model loaded, or BYOK key saved).
2. Click any report row to open the dialog.
3. Switch to the AI summary tab.
4. Click Generate. The model streams a 4-section plain-English briefing (what it's about, key findings, recommendations, why it matters).
5. The summary is cached in your browser — opening that report again later shows it instantly. Click Regenerate for a fresh attempt.

8. Asking questions about a report

Ask tab — chat input scoped to the open report.

1. From the dialog, switch to the Ask tab.
2. Type a question. Press Enter or click Send.
3. The model gets the report's full text (truncated to fit the context window) plus your cached summary if one exists. It answers from the report only.
4. Each new question + answer is appended to the thread for the duration of the dialog. Closing the dialog resets the thread.

9. Web-search enrichment for Ask

Settings → Web search section.

Optional. When configured, you get a 🌐 button next to Send in the Ask tab. Clicking it does a web search first and feeds the top results into the prompt alongside the report text. Useful for "is there recent news on this?" or "has the government acted on this since the report?" style follow-ups.

1. Settings → Web search section.
2. Pick a provider. For Tavily / Brave, paste your API key. For SearXNG, paste the instance URL (e.g. https://searx.example.org).
3. Save. The 🌐 button appears in the Ask tab.

10. Exporting metadata and summaries

Two buttons live in the toolbar above the report list:

• ⬇ Metadata (CSV) — exports the currently filtered reports as CSV. 13 columns including has_text and has_summary. Useful for dropping into Excel / a notebook for analysis.
• ⬇ Summaries (MD) — exports every AI summary you've generated as a single Markdown file, with each report headed by title + committee + LS + report number.

Both downloads happen client-side — nothing leaves your machine.

11. Deep search (optional)

By default, search hits report titles. The instant you open a report's Full text tab, that body becomes searchable too — and stays cached locally forever. If you want to search across the entire corpus' body content without opening reports one-by-one, enable Deep search in Settings — independently per corpus.

1. Open Settings → scroll to the per-corpus search section for the corpus you want to deep-search.
2. Tick "Enable full-text search across <corpus>". The estimate line tells you how much will download.
3. Save. The app starts fetching every extracted text in the background (throttled). A small status appears below the filter row while it indexes.
4. Once done, search hits body content for that corpus's full backfill. Cached locally — return visits use zero new bandwidth for it.

12. Privacy & what's cached locally

SansadSaar has no server. Everything you generate stays in your browser:

• API keys — localStorage, never sent except to the provider you chose.
• Generated summaries — IndexedDB store summaries. Persistent across sessions. Visible in DevTools → Application → IndexedDB → sansadlocal (legacy DB name, kept so existing data carries over).
• Extracted PDF text — IndexedDB store texts. Once fetched, it's there for life. On return visits, only newly-published reports get pulled from the mirror.
• Model weights — browser Cache Storage. ~1.5–4.9 GB depending on the model.
• Settings — localStorage key sansadlocal.settings.v1 (legacy key, preserved across the rename).

What leaves your browser:

• Fetches to the data-mirror origins, all served as static assets from Cloudflare Workers:
  • sansadsaar-data.naklitechie.com — DRSC, CAG, FC, LC, Bills (the document-corpus mirror)
  • sansadsaar-proceedings.naklitechie.com — Debates + Questions
  • sansadsaar-gazettes.naklitechie.com — Central Gazette
  No tracking, no analytics.
• Click-through to archive.org when you open a Gazette's "View PDF" — that's a normal external link, the app doesn't proxy.
• Fetches to huggingface.co — model weights, only on first load of a given model.
• BYOK requests go directly to your chosen provider. Your key isn't proxied anywhere.
• Search-enrichment requests go directly to your chosen search provider.

What never happens: no analytics, no accounts, no telemetry, no server logs (no server). The page is a static index.html served from Cloudflare Workers, on the custom domain.

13. Costs & limits — what's free, what isn't

Three layers of cost. The first is always free; the other two depend on your choice.

Layer 1 — SansadSaar infrastructure

Layer 2 — Local AI inference

Layer 3 — BYOK API providers (optional)

Layer 3b — Web-search enrichment (optional)

14. Troubleshooting

The local model won't load

• Check the AI status pill — if it says "No WebGPU", the browser doesn't support WebGPU. Use Chrome 113+, Edge 113+, Brave, or Firefox 130+.
• If it shows "Worker error", open DevTools console for the underlying message (network failure, OOM, model file corruption). The "Clear cache" button in Settings forces a re-download.

Generation is slow

• Local AI on a CPU-only machine is slow — WebGPU on a recent GPU is needed for usable speed. Check chrome://gpu in Chrome.
• Long reports get truncated to fit the model's context window (8K for Gemma 4 E2B, 32K for Bonsai 1.7B/4B, 64K for Bonsai 8B). For very long reports, the larger Bonsai models or a BYOK provider produce better summaries.

Ollama returns CORS errors

Ollama's HTTP API doesn't allow cross-origin requests by default. Set OLLAMA_ORIGINS when starting Ollama:

OLLAMA_ORIGINS="https://sansadsaar.naklitechie.com" ollama serve

If you opened the app via http://localhost:8000 for local dev, set OLLAMA_ORIGINS=http://localhost:8000 instead.

The mirror is missing data

Each corpus has its own GitHub Action on its own schedule — DRSC daily, CAG / FC / LC / Bills daily, Debates 2-hourly, Questions hourly, Gazettes hourly. A new upstream record typically appears in SansadSaar within that corpus's next cycle. On a fresh visit, the latest data loads automatically; if you've linked Save-to-Disk, the staleness check picks up updates and re-syncs in the background.

Where do credits live?

Help → Credits tab. Open from the ? button in the header.

Built on top of ParliamentWatch by Pranay Kotasthane — the original DRSC scraper, committee config, and core idea are his. The Gazettes corpus shape is inspired by egazette by Sushant Sinha (independently implemented; no code copied). SansadSaar repackages all of this with on-device AI and the additional corpora. Full credit list in the Help → Credits tab.