Skip to main content

Agent-retrieval layer on top of the Cairn entity engine — adaptive gate, signal orchestration (lexical/semantic/graph), and minimal context assembly. Zero generative-LLM on the hot path.

Project description

cairn-retrieval

The agent-retrieval layer on top of the cairn-engine entity engine.

The three retrieval primitives above the entity map — all deterministic, all generative-LLM-free on the default path:

  • A · Adaptive gate (gate) — two-stage, LLM-free decision before any signal runs: deterministic bypass (temporal markers, multi-entity relational, private entities) then a CA-RAG utility cost function. strategy=none (skip retrieval) is a first-class, reachable outcome — selective retrieval beats always-on empirically.
  • C · Signal orchestrationLexicalIndex (pure-Python BM25) + scan (exact/regex), opt-in SemanticIndex (caller-supplied embedder; no model SDK imported), graph traversal via cairn-engine; ranked lists merged with fuse (Reciprocal Rank Fusion, k=1, rank-only — no score normalization).
  • E · Context assembler (assemble) — budget → MMR select → cosine dedup → fold ordering (strongest evidence at the edges) → provenance-tagged emit with a trace manifest. Never re-embeds; zero LLM.

RetrievalEngine.retrieve(query, *, budget) ties them end-to-end: resolve → gate → orchestrate → fuse → assemble. A query the gate routes to none returns no context at all — the skip is the point.

Also here: rank_links — TF-IDF-cosine document↔document link recommendation over shared canonical concepts — and suggest_connections — cold-start link prediction: rank the existing entities a NEW candidate node should connect to (content-first, seed-optional structural signal, full provenance).

Status: feature-complete, 0.x. Stays 0.x until the benchmark phase (OP-31 eval vs. agentic-grep / vector-RAG / Graphiti) proves the quality claims. Depends only on cairn-engine (which is itself zero-dependency). See the roadmap and intents/ for the constraint-tested design of each module.

Install

pip install cairn-retrieval   # pulls cairn-engine automatically

License

MIT — Copyright (c) 2026 Leo Celis

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

cairn_retrieval-0.1.0.tar.gz (48.0 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

cairn_retrieval-0.1.0-py3-none-any.whl (21.4 kB view details)

Uploaded Python 3

File details

Details for the file cairn_retrieval-0.1.0.tar.gz.

File metadata

  • Download URL: cairn_retrieval-0.1.0.tar.gz
  • Upload date:
  • Size: 48.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.11.0

File hashes

Hashes for cairn_retrieval-0.1.0.tar.gz
Algorithm Hash digest
SHA256 01c6c46e32ea0c6aaec94afa88e30efab1528ea99f6c0180cd1f576882fca530
MD5 86d247231b0deae19628dde8e589159a
BLAKE2b-256 c7074afa59630be6363db010852ccf5ae4f6436f0f13e60252c252c493878198

See more details on using hashes here.

File details

Details for the file cairn_retrieval-0.1.0-py3-none-any.whl.

File metadata

File hashes

Hashes for cairn_retrieval-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 22401b0ca36637df55ee214335553526b2eddcba26bfccc97314f58c7fc6de2e
MD5 266109ac61ced100aa2df5cea1a8d23f
BLAKE2b-256 d381104d8fa843405bf5fbffdc5bfdcde4e6b15e4187c694ff8060a04f0814d1

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page