pvcouncil — Prairie Village City Council document search

As a sitting councilor, a group of questions would routinely come up during meetings and discussions with constituents: "When did this item pass?" "Who voted for it?" "What were the discussions around it, if any?"

I made this tool to process and document council packets going back to 2015 (the earliest ones still available) so I could get some historical context into past council thought and decision-making processes. The technical bits follow.

Semantic and keyword search, plus cited RAG Q&A, over Prairie Village, Kansas city council packets and minutes (2015–present). PDFs are chunked and embedded into a local ChromaDB index with Voyage AI embeddings; pvc ask synthesizes answers with Claude, citing the source packet, page, and meeting date.

Documents are fetched from the city's public CivicClerk portal API. Scanned/image-only PDFs are OCR'd with Tesseract.

Setup

Requires Python 3.12+, uv, and optionally Tesseract (brew install tesseract) for OCR of scanned documents.

cp .env.example .env   # add your Voyage AI and Anthropic API keys
uv sync

Usage

uv run pvc download    # fetch new packets/minutes from CivicClerk
uv run pvc index       # chunk + embed new documents (incremental)

uv run pvc search "stormwater drainage"
uv run pvc search "ordinance 2428" --keyword
uv run pvc search "housing density" --type packet --after 2023-01-01

uv run pvc ask "What has the council decided about the community center?"

uv run pvc stats

Voting records

pvc extract-votes runs every minutes document through Claude with a JSON schema, producing votes.json: each formal motion, its mover/seconder, result, tally, and — where the minutes explicitly record them (roll calls, named dissents) — who voted yes/no/abstain. Individual votes are never inferred from attendance. Extraction is incremental; pass --batch to use the Message Batches API (50% cheaper) for large re-extractions.

uv run pvc extract-votes                  # extract new minutes (incremental)
uv run pvc extract-votes --batch --force  # full re-extraction via Batches API

uv run pvc votes --result failed --after 2024-01-01
uv run pvc votes --member Reddell --side no
uv run pvc votes --search "community center" --summary
uv run pvc votes --markdown council-votes.md   # regenerate the full report

The full voting record is checked in as council-votes.md — every recorded vote from April 2015 onward, grouped by meeting.

An MCP server (pvcouncil.mcp_server) exposes search, ask, votes, list_documents, and stats as tools for MCP clients.

Data

pv_council_docs/ (source PDFs) and chroma_db/ (the vector index) are not tracked in git. Rebuild them from scratch with pvc download && pvc index (embedding the full corpus costs roughly $0.20 in Voyage API usage).

License

AGPL-3.0-or-later