Distill

A modular research knowledge and briefing engine.

Distill watches sources — arXiv, research lab blogs — scores incoming signals against a durable topic wiki, and generates strategic briefings tailored to a target audience. Both the research topic and the audience are configuration, not code.

The first example covers emerging data advantages in AI, briefed for technical decision-makers. Swap the topic config and audience profile to cover any research domain and reader.

How it works

1. Source adapters fetch and normalise items from arXiv RSS feeds and lab blogs
2. Scoring pipeline runs two passes: a cheap model filters on abstract relevance, an expensive model scores the full signal
3. Wiki updater classifies signals as replication, adjacent, or new; writes to theme files, entities, and timeline accordingly
4. Briefing generator produces a structured briefing for a target audience profile

Storage is file-based: Markdown for narrative (themes, overview), JSON for structured data and belief state (entities, timeline, hypotheses, evidence). No database.

Setup

git clone https://github.com/your-username/distill
cd distill
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env
# edit .env and add your GEMINI_API_KEY

Running tests

PYTHONPATH=src pytest tests/

Tests are offline by default — the Gemini client is mocked. No API key required for pytest.

Bootstrapping a new topic

1. Write a deep-research dossier for your topic and save it to data/research_topics/{topic_id}/dossiers/bootstrap_{date}.md
2. Run the seeder:

PYTHONPATH=src python -m topics.bootstrap.seeder \
    --dossier data/research_topics/{topic_id}/dossiers/bootstrap_{date}.md

3. Review the generated wiki under data/research_topics/{topic_id}/

See docs/research_briefing_architecture.md for the full architecture and docs/15_knowledge_management_briefing_engine.md for the implementation plan.

Project layout

src/          Python package root (PYTHONPATH=src)
tests/        pytest suite with offline fixtures
data/         Topics, audiences, credibility data
docs/         Architecture, specs, diagrams

Contributing

Read AGENTS.md for project conventions and CODING_PRINCIPLES.md for code style.