Drop documents here that describe what you are actually working on, and the scorer will judge every paper and article against them.
This is usually the highest-leverage tuning available. A two-page concept note
tells the scorer more about what matters to you than any number of query strings
in registry/sources.yaml — those decide what gets collected, these decide
what counts as relevant.
- A concept note, proposal, or grant abstract
- A thesis chapter outline, or a description of the argument you're building
- A short "what I care about and what I don't" note — exclusions help as much as inclusions
- A reading list with a line on why each item mattered
- Anything confidential. Everything in this folder is committed to the repository, and the repository may be public. Unpublished results, personal data, and anything under embargo do not belong here.
- Whole PDFs of papers. Only
.md,.txt,.markdownand.rstare read — a PDF read as bytes would be noise, and adding a parser would cost a dependency this project deliberately avoids. - Anything enormous. The combined text is truncated at
context.max_chars(24,000 by default), so a large drop silently loses its tail. Two or three focused pages beat twenty unfocused ones.
Every file is read in filename order, concatenated, and appended to the system prompt used for scoring. Each run prints what it loaded:
context: 2 reference documents, 8431 chars
LLM scoring only. The free keyword scorer doesn't read these files — it
matches the term lists in registry/sources.yaml. If you're running without an
API key, set scoring.provider to a free option (groq, gemini) to get any
benefit from what you put here. See the scoring table in the README.
Turn it off with context.enabled: false, or point it elsewhere with
context.dir.
This README is skipped automatically — it's the folder's own documentation, not a research note. Delete it if you like; the folder works either way.