Deep research
Alpha TestingLocal-first Deep Research System
Local-first deep research agent — open source, alpha testing.
Overview
Local-first deep research agent — open source, alpha testing.
Problem
In programme and architecture reviews, keyword search often misses semantically relevant passages in technical documents — I need hits with explainable relevance, not just a similarity score.
Why I built it
Ada answers questions over my knowledge. DeepDiver explores how well embeddings represent technical domain language — and whether hybrid search works better in practice than purely semantic.
Who benefits
Technical users and researchers searching semantically over programme documentation, papers, or internal specs — primarily myself when testing RAG and search patterns.
Current status
Alpha testing — trustworthy, explainable research workflows.
Architecture
Current implementation — based on what is built today.
Technology stack
Python · Sentence Transformers · FAISS · FastAPI · React
Lessons learned
- Hybrid search (vector + keyword) is often more practical than purely semantic
- UI must surface uncertainty in similarity scores
- Research ops starts with reproducible index pipelines
Future direction
- Evaluation framework for domain-language hits
- MCP tool server for semantic search as an agent tool
- Explainable relevance scores in the UI