VR-ETL

Vectorless, agentic QA over the Cyber Ireland 2022 report. ETL ingests the PDF into a PageIndex tree + node map; a single /chat endpoint orchestrates retrieval, math tooling, and streaming answers with citations.

Query test logs: see test_queries directory for the three evaluation scenarios.

Tech Stack

• FastAPI backend with LangGraph orchestration
• PageIndex for document ingestion and hierarchical retrieval
• DeepSeek (OpenAI-compatible) for reasoning + tool calls
• Sympy math tool (strict function-calling protocol)
• Static frontend (vanilla HTML/JS) with streamed NDJSON and Mermaid rendering

Documentation

• Setup — environment, env vars
• ETL — pipeline steps, artifacts
• Backend — endpoints, prompts, flow
• Frontend — UI usage
• Architecture — high-level flow and justification slots

Quick Start

pip install -r requirements.txt
cp .env.example .env  # fill PAGEINDEX_API_KEY, API_KEY, BASE_URL, MODELS
python run_etl.py     # runs the ETL piepline and generates doc_id, tree, node_map
uvicorn src.backend.api:app --reload # starts backend endpoint
# open src/frontend/index.html (or serve via python -m http.server -d src/frontend 5500)

Notes

• Data artifacts (data/processed/*) and logs are git-ignored; regenerate via ETL.
• /chat streams NDJSON events: meta → tool (if math) → reason → answer → done.