docq

Turn a source text into a small, portable, cited corpus you can search by free-text query and metadata — and get back the source's actual passages, with citations, instead of a paraphrase. Retrieval is hybrid: FTS5/BM25 plus an optional local-embedding channel, fused by reciprocal rank fusion, with each hit tagged by which channels ranked it (two agreeing channels = a hit you can trust).

docq splits into a corpus-agnostic engine (src/docq/) and per-book instances (corpora/<name>/). The first — and so far only — instance is John Ousterhout's A Philosophy of Software Design (APOSD): a deep, searchable backing for a software-design skill, surfacing the book's own words and examples on demand.

Status: early. It builds the full APOSD corpus end-to-end and hybrid retrieval scores hit@5=0.94 / hit@1=0.71 / MRR=0.80 on a 31-case eval, but real-world usefulness hasn't been battle-tested. Treat it as a working prototype, not a finished product.

How it works

Two phases (plus an optional embed pass), one portable artifact:

• Build (offline, once, needs a model): PyMuPDF font-aware parse → deterministic unit segmentation (verbatim text is fixed here, never LLM-rewritten; code blocks travel with the prose that introduces them) → per-unit LLM enrichment (retrieval metadata: a context line, symptom-phrased questions, key terms) → a single SQLite/FTS5 .db. Checkpointed, resumable, and concurrent.
• Embed (offline, optional, ~16s): docq embed writes per-unit vectors — a metadata gist, one vector per generated question, chunked verbatim text — into the same .db, via a local Ollama embedding model (embeddinggemma).
• Query (repeated): BM25 and vector max-similarity as independent channels, fused with reciprocal rank fusion; every hit shows its per-channel ranks (via lex#1+sem#2). Stdlib-only packages — lexical mode needs literally nothing installed or running; hybrid needs only the local Ollama service, and degrades back to lexical without it.

Boundaries and verbatim text are deterministic; the LLM only classifies and generates retrieval fields, so what comes back is always the source's own words.

Quickstart

# Once published, install persistently or run ephemerally:
uv tool install docq
uvx docq --help

# Tests (no model, no corpus needed):
uv run --extra build pytest -q

# Build a corpus (needs the source PDF locally + an Ollama model — neither is included):
uv run --extra build docq build --model <model> --workers 8 --build-dir build/<model>

# Embed it for hybrid retrieval (stdlib-only; needs local Ollama + embeddinggemma):
uv run docq embed

# Retrieve (stdlib-only — no extras; hybrid when embedded, lexical otherwise):
uv run docq retrieve "should I make this API general purpose" -k 3 \
    [--principle general-purpose] [--type red_flag] [--json] [--mode lexical|hybrid|semantic]

# Eval — hit@k / hit@1 / MRR over corpora/<name>/cases.yaml:
uv run --extra build docq eval --mode hybrid

This checkout includes .docq/config.json, which supplies its APOSD paths. An installed copy instead reads $XDG_CONFIG_HOME/docq/config.json (or ~/.config/docq/config.json) and then <git-root>/.docq/config.json; project values override user values, and CLI flags override both. An explicit --config FILE uses only that file plus CLI flags. Run docq config to see the effective values and files considered.

The SQLite corpus is deliberately not included in the wheel. A typical personal configuration is:

{
  "db": "~/.local/share/docq/corpora/aposd.db",
  "retrieve": {
    "rerank-model": "gemma4:e2b"
  }
}

Layout

src/docq/        engine — corpus-agnostic: parse, segment, enrich, store, cli, taxonomy
corpora/aposd/    APOSD instance — profile, taxonomy, enrichment prompt, eval cases
docs/             design spec, implementation plans, session notes
tests/

Provenance & license

The docq engine and tooling are MIT-licensed (see LICENSE).

The software-design-philosophy skill (.claude/skills/) and the APOSD taxonomy (corpora/aposd/taxonomy.yaml) are distillations of John Ousterhout's A Philosophy of Software Design, kept here as the development input that seeded the corpus — derived summaries, not original work. The book itself (the source PDF) and any built corpus database are not distributed (they're gitignored); building a corpus requires your own copy of the source.