AI Document Search for UK Parish Councils

UK parish councils publish thousands of documents: meeting minutes, agendas, budgets, audit reports, planning notices. But there was no way to search across them.

Council clerks and residents who needed to find a specific decision, figure, or outcome had to open individual PDFs one by one. A question like "what did the council decide about the car park in 2024?" meant hours of manual digging through dozens of files.

No cross-document search existed. No way to ask a question in plain English and get an answer with a source.

I built an AI-powered search system that ingests parish-council documents, embeds them using Google Gemini, and answers plain-English questions with exact citations pulled from the source PDFs.

Document ingestion pipeline — automated scrapers pull new documents from council websites within hours of publication

OCR pipeline — handles scanned PDFs that aren't machine-readable

Vector embeddings + hybrid search — PostgreSQL with pgvector stores document embeddings; combines semantic search with keyword matching for better recall

Cross-encoder reranking — sentence-transformers rerank results to surface the most relevant passages

Citation extraction — every answer includes the exact source document and passage, so users can verify

Plain-English interface — non-technical users (clerks, residents) can query complex topics with no training

Streamlit admin panel — for managing documents, monitoring ingestion, and system configuration

Automated updates — GitHub Actions CI/CD keeps the system current as new documents are published

Frontend: Next.js

Backend: FastAPI + SQLAlchemy + Alembic

AI/ML: Google Gemini embeddings, cross-encoder reranking (sentence-transformers)

Database: PostgreSQL + pgvector for vector storage and hybrid search

Infrastructure: Nginx reverse proxy, GitHub Actions CI/CD

~10,000 documents indexed across multiple UK parish councils

Questions that previously required opening dozens of PDFs now answered in seconds with exact citations

Automated scrapers keep the corpus current — new documents appear within hours of council publication

Non-technical users query complex topics in plain English with zero training