pinakes

A portable, agent-first knowledge base. One directory = one KB.

The Pinakes were Callimachus's catalogue of the Library of Alexandria — the first known index of a body of knowledge.

📖 Guide · CLI · Manifest · Design · What ships today

---

The idea

A knowledge base is a plain directory you can read, edit, diff, commit and hand to someone:

my-kb/
├── pinakes.toml              # manifest: sources, models, chunking, budget
├── docs/                     # SOURCE OF TRUTH — your files, unmodified
│   ├── paper.pdf
│   ├── paper.pdf.pnk.yaml    # sidecar: stable ID, tags, links, provenance
│   ├── notes.md
│   └── notes.md.pnk.yaml
└── .pinakes/                 # generated, disposable, gitignored
    └── index.db              # SQLite: chunks, FTS5, vectors, links

Your documents and their metadata are the truth. The index is derived state that can always be rebuilt. That split is what makes a KB both a reproducible recipe and a directory you can move.

What makes it different

It costs nothing to run. Retrieval is BM25 (SQLite FTS5) + local embeddings + local reranking, fused and scored entirely on your CPU. No API key is needed to search, and re-indexing is free — so there is never a cost reason not to improve your chunking or swap your embedding model. That the free path stays free is enforced by a CI gate, not by a promise.

Reasoning is the caller's, not the KB's. The MCP tools return ranked, cited evidence. pinakes_search → pinakes_get → pinakes_search is a plan-retrieve-read-refine loop, and your agent already runs it in its own context. Multi-hop reasoning falls out of composable tools rather than a second agent framework.

Money is opt-in and bounded by design. Every paid path is an explicit, enumerated entry point, and a pre-call reservation makes a hard cap a real ceiling rather than an after-the-fact report. See what is actually built.

KBs link to each other. Sidecars carry pnk://<kb-ulid>/<doc-ulid> references, so links survive renames, moves, and being shared with someone else. The addressing ships today because IDs cannot be retrofitted; authoring and traversing those links is the links release.

Its limits are published, not hidden. No vector tier is sublinear; cross-KB answers will be capped by how well your KBs are linked; and the confidence heuristic's measured false-confidence rate is 0.25 — one no-answer question in four still gets a confident answer. A heuristic whose cost is unmeasured is worse than one whose cost is known.

Quickstart

uv add "pinakes[st]"                  # default backend
uv add "pinakes[light]"               # fastembed, no torch
uv add "pinakes[light,pdf]"           # + PDF ingest, free and local
uv add "pinakes[light,pdf,claude]"    # + the opt-in paid extractor for scanned PDFs

[claude] installs a path that can spend money, and nothing spends without you asking: the default extractor is free, and reaching the paid one takes --extract=claude-vision (or a manifest key) and a real API key in the environment. When you do ask, the run is priced before the first call and refused if it would breach any of the three [budget] caps — per_operation_eur, daily_eur or monthly_eur. Raising one and hitting the next is the discovery path those caps exist to prevent, so a refusal names every window that binds, not just the first. pnk budget reports what has been spent.

pnk init my-kb                        # stamp a KB
pnk sync                              # index what changed (git-hook friendly)
pnk search "hybrid retrieval"         # free: BM25 + vector + rerank
pnk doctor                            # environment, coherence, orphans, link coverage

uvx --from "pinakes[st]" pnk serve    # MCP server, nothing installed

⚠️ Two things pnk init cannot know, each needing one manifest edit: on a [light] install set provider = "fastembed", and to index PDFs add "**/*.pdf" to [sources] include. Both are in the Guide.

→ Full guide — PDFs, filters, calibration, git hooks, MCP setup, troubleshooting.

Development

make install    # sync the dev environment (the light extra — CI's minimum leg)
make check      # every gate, stopping at the first failure — run before every commit
make demo       # index the synthetic demo KB
make eval       # golden-set evaluation against the recorded baseline
make corpus     # regenerate the synthetic PDF corpus in place
make pdf-eval   # extraction-quality baseline + floor-drift check (needs [pdf])
make budget     # the demo KB's spend ledger (free: it only reads)
make help       # all targets

Every target wraps the command CI actually runs, so green locally means green on the runner. Note that make check formats Python inside Markdown fences too — a docs-only change can fail it.

Conventions and the increment workflow are in CLAUDE.md; how the docs are organised — and which file to edit when you land a feature — is in docs/README.md. docs/VERIFICATION.md maps every promise this project makes to the test that holds it, and a test asserts every one of those tests exists.

Your data stays yours

This repository contains the engine only. Real knowledge bases live outside it. The sole KB here is a small synthetic corpus used for tests and retrieval benchmarking, and the only other committed content is tests/pdf-corpus/ — PDFs generated from scratch by a committed script to exercise hard extraction cases. Neither was harvested from anywhere; no real-world document is committed here.

pnk init ships a .gitignore covering .pinakes/, so your index — and your spend ledger — never leaves your machine. Note that publishing a KB repo publishes docs/ and every sidecar: titles, tags and provenance URLs included.

Licence

Apache-2.0.