Chat with My Research

Follow the evidence

1. Plan focused queries, preserving the original question.
2. Retrieve relevant passages from the local SQLite FTS index.
3. Optionally rerank passages with Cohere.
4. Answer with source citations—or show an extractive preview without model credentials.

Compare planning strategies

Compare MCTS, an admissible A* retrieval-cost variant, and beam search against Self-Ask, least-to-most, linear, and ReAct-style baselines. Inspect the bounded search tree and plot labeled multi-hop source accuracy against retrieval calls.

The A* heuristic uses a zero lower bound on remaining non-negative call cost; normalized retrieval confidence breaks ties. Confidence and the default value function are lexical estimates, not calibrated probabilities or answer-correctness guarantees.

Try the local demo

cd chat-with-my-research
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
research-chat ingest data/demo-corpus
streamlit run dashboard.py

The included notes are synthetic project fixtures, not representations of my papers, talks, or personal views. Add your own authorized corpus locally; the index is git-ignored by default. Without an API key, the demo shows extractive passages and makes no model-provider calls.

Run the strategy comparison from the dashboard, or useresearch-chat compare --dataset samples/evaluation.jsonl. The sample scores are synthetic demonstration data, not published benchmark results.

Example questions

• What does the sample note say about agent autonomy?
• Compare evaluation to observability.
• Which sample passage discusses tracing answers to sources?

Source-level metrics include recall@k, precision@k, reciprocal rank, citation precision, unsupported citations, and optional required-term coverage.

Back to profile