Vimarsh (विमर्श, "deliberation") publishes neutral, plain-language analysis of bills in the Parliament of India. A daily GitHub Actions job reads the official bill listings on sansad.in, downloads the latest official text of each new or amended bill, extracts the text locally, asks a language model for a structured analysis, and commits the result to this repository. A second workflow builds a static site, an RSS feed and a JSON index, and publishes them on GitHub Pages.
In keeping with Pardarsh, every analysis is attributed and checkable: each file records the official document URL, the SHA-256 of the document and of the extracted text, the provider, the requested and served model, the response ID, generation settings (temperature, thinking level), token usage including thinking tokens, latency, attempts and any fallback, the prompt and schema versions with a hash of both (the prompt itself is published on the site's Method page), software versions, a link to the GitHub Actions run, and the model's own summary of its reasoning when the provider returns one.
sansad.in listings ──► sources.py ──► data/manifest.json (state, in git)
│ pending bills (max 15 per run)
▼
official PDF ──► extract.py ──► providers/router.py ──► Gemini ─(429/503 ×3)─► Groq
│
▼
content/bills/<year>/<bill-id>.md (YAML frontmatter + Markdown)
│ deploy workflow
▼
public/ (index, bill pages, feed.xml, bills.json) ──► GitHub Pages
- Sources. Only sansad.in.
api_rs/legislation/getBillslists bills from both houses;api_ls/legislation/getBillsis also read and skipped with a warning if it fails. For each bill the most recent text is used (passed by both houses, passed by one house, or as introduced). - State.
data/manifest.jsontracks each bill'sstatus(pending,processed,failed,skipped_oversized),version_hash(SHA-256 of the document),source_url, attempts, last error andlast_updated. A bill is analysed again only when a new document is published and its bytes differ, so reruns cost nothing. When only a bill's listing details change (for example its status goes from Passed to Assented), its page is updated without calling the model again. - Providers.
LLMProvider.analyze_bill(text, schema) -> AnalysisResultinsrc/vimarsh/providers/base.py, with adapters for Gemini (google-genai) and Groq (groq). Output is validated with Pydantic. Transient errors (429, 503) are retried three times with exponential backoff, then the fallback provider is tried. If every provider fails, the bill stayspendingand is retried on the next run; the run never crashes. - Guardrails. Calls are spaced by
REQUEST_DELAY_SECONDS. Tokens are estimated locally (about four characters per token) before each call. Text over a provider's input budget is cut to its beginning and end (clauses, Statement of Objects and Reasons, Financial Memorandum) and the result is markedpartial_analysis: true.
Set these as repository variables (Settings → Secrets and variables → Actions → Variables); all are optional.
| Variable | Default | Meaning |
|---|---|---|
PRIMARY_PROVIDER |
gemini |
First provider (gemini or groq) |
FALLBACK_PROVIDER |
groq |
Second provider, or none |
GEMINI_MODEL |
gemini-3.8-flash |
Gemini model id |
GROQ_MODEL |
llama-3.3-70b-versatile |
Groq model id |
GEMINI_THINKING_LEVEL |
low |
Gemini thinking level; none to omit it |
REQUEST_TIMEOUT_SECONDS |
300 |
Give up on a model call after this long (then retry or fall back); also Gemini's server-side deadline |
MAX_CONSECUTIVE_FAILURES |
2 |
Stop a run after this many bills in a row fail on every provider |
REQUEST_DELAY_SECONDS |
12 |
Minimum gap between API calls |
MAX_BILLS_PER_RUN |
15 |
Bills analysed per run; the rest stay pending |
MAX_RETRIES |
3 |
Retries on 429/503 before falling back |
SINCE_DATE |
2026-01-01 |
Ignore bills introduced earlier |
GEMINI_MAX_INPUT_TOKENS |
120000 |
Truncate above this for Gemini |
GROQ_MAX_INPUT_TOKENS |
8000 |
Truncate above this for Groq |
SITE_URL |
https://vimarsh.pardarsh.9th.fun/ |
Absolute URL used in the feed |
API keys go only in secrets: GEMINI_API_KEY and GROQ_API_KEY. They are
never written to the repository or the site. Only public bill texts are sent
to the providers.
python -m venv .venv.venv/bin/pip install -e ".[dev]"GEMINI_API_KEY=... MAX_BILLS_PER_RUN=1 vimarsh runvimarsh build --out publicOpen public/index.html in a browser.
- In Settings → Pages, set the source to GitHub Actions and the custom
domain to
vimarsh.pardarsh.9th.fun(theCNAMEfile), with a DNSCNAMErecord pointingvimarsh.pardarshatpardarsh.github.io. Tick Enforce HTTPS once the certificate is issued. - Add the
GEMINI_API_KEY(and optionallyGROQ_API_KEY) secrets. - Run Analyze bills (daily) once by hand, or wait for its daily run at 02:30 UTC (08:00 IST).
There are three workflows:
- Analyze bills (daily) (
analyze.yml): every day, and by hand, analyses up toMAX_BILLS_PER_RUNpending bills, newest first, introduced on or afterSINCE_DATE. - Analyze specific bills (
analyze-bills.yml): by hand only. Enter bill IDs (the file name undercontent/bills/, for examplels-2026-97), tick force to redo an already-analysed bill, and set since_date for bills introduced beforeSINCE_DATE. Locally:vimarsh run --bill ls-2026-97 --force. - Deploy site (
deploy.yml): runs after either of the above and on pushes tomaster; it never commits, so it cannot trigger an analysis loop.
Both analysis workflows run the same job (_analyze.yml), never at the same
time, and commit only data/ and content/: at most once per run, and only
when at least one bill was analysed.
pytestpre-commit run --all-filesMIT; see LICENSE.txt. Bill texts are published by the Parliament of India.