RE-XTRACT

Unanimity-gated document extraction with calibrated abstention, for Indian real-estate regulatory (RERA) filings, plus the AAAI-27 demonstration-track paper describing it.

Every routing and extraction decision passes three gates over a self-hosted open-weight VLM:

1. Voting — M=5 votes from one model must agree exactly, retried up to N-1 times
2. Verification — a frontier model must independently reproduce the unanimous answer
3. Escalation and abstention — K-1 further frontier calls, and if those do not agree the pipeline writes nothing rather than guessing

Layout

demo/     web UI that runs the pipeline live and streams its decision trace
paper/    AAAI-27 demo paper (LaTeX, AAAI AuthorKit27 style)
agents/   classifier and extractor, each wrapped in the voting stack
tools/    voting, escalation, PDF rendering, date normalisation, inference backends
db/       Databricks read path
config.py all tunable parameters

Setup

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env      # then fill in the values

The open-weight model is served over Ollama at the host set in config.py (OLLAMA_API_URL). That host must be reachable from wherever you run the demo.

Running the demo

uvicorn demo.server:app --port 8000

Open http://localhost:8000, pick one of the six preloaded projects, and the pipeline runs on it live. The centre panel streams the pipeline's own log: every vote as it lands, the validator's verdict, each round transition, and the record written or the abstention.

The demo never writes

It reads document URLs from Databricks through get_document_urls() and nothing else. It deliberately does not import graph.workflow from the production system, because that module builds a storage backend at import time and writes in two places; its read path also issues CREATE TABLE IF NOT EXISTS. It also performs no MD5 or perceptual-hash deduplication, so every document in a project is processed and shown.

Building the paper

cd paper && latexmk -pdf main.tex

Produces two pages of content plus one page of references, per the AAAI-27 call. paper/README.md records what is still outstanding before submission.

paper/reference/ holds accepted AAAI-26 demo papers kept for style comparison. They are other authors' work and are gitignored.