oiax 0.1.4

oiax — semantic policy routing for agent fleets. Delivers the right governance context to the right agent at the right turn, by meaning.

oiax

oiax — semantic policy routing for agent fleets. The tiller that delivers the right governance context to the right agent at the right turn, by meaning.

oiax routes a free-text prompt against a governance corpus and delivers the policies that bear on that turn — by meaning, before the agent decides, at ~6ms with no network call.

The retrieval design is for normative text (policies, coding standards, ADRs, compliance rules), not general knowledge. Four decisions make it correct:

1. Whole-document delivery, never chunks. A rule and its carve-out are semantically distant but logically inseparable.
2. Precision over recall, asymmetric errors. A miss degrades to the status quo; a false positive actively degrades the layer.
3. Surface names, never rules. Matched terms make a bad match dismissible at a glance.
4. Runtime-agnostic core, harness-specific adapters. The router returns structured hits; each harness gets its own thin delivery layer.

Installation

pip install oiax

Requires Python ≥ 3.11. On first use, a ~90MB ONNX embedding model downloads and caches locally. Subsequent routes are ~6ms.

Quick start

Route a prompt

from oiax import build_index, route
from oiax.corpus import PolicyDirCorpus

# Load from a directory of markdown files, each carrying an **Agent-trigger:**
# line (see "Corpus format" below)
corpus = PolicyDirCorpus("./policies/")
index = build_index(corpus)

# Route a prompt — at most two hits, ranked by reciprocal-rank fusion
hits = route("How do I deploy to production?", index)
for hit in hits:
    print(f"{hit.name} ({hit.score:.2f}): {', '.join(hit.why)}")

Route with query expansions

import json

expansions = json.load(open("./routing-expansions.json"))
index = build_index(corpus, expansions=expansions)
hits = route("help me merge my PR", index)

Use a custom corpus

from oiax.corpus import Document

class MyCorpus:
    def documents(self):
        yield Document(
            name="deploy-policy",
            trigger_line="deploying to production",
            body="Always run the test suite before deploying...",
        )

hits = route("deploy to prod", build_index(MyCorpus()))

Corpus format

Policy files are markdown with an **Agent-trigger:** header — a one-line statement of what the document governs. This is used for both lexical matching (TF-IDF) and semantic matching (embeddings).

# My deploy policy

**Agent-trigger:** deploying the application to production, CI/CD configuration

Always run the test suite before deploying. Never deploy on Friday.

The PolicyDirCorpus loader reads all *.md files in a directory. The filename (without .md) becomes the document name returned in route hits.

Claude Code integration

oiax.adapters.claude_code is a UserPromptSubmit hook adapter. Register it in ~/.claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [{
          "type": "command",
          "command": "python3 -m oiax.adapters.claude_code /path/to/policies/ --expansions /path/to/expansions.json",
          "timeout": 8
        }]
      }
    ]
  }
}

On every prompt, the adapter routes the prompt text against the policy corpus and injects a context paragraph naming the policies that may apply — with the matched terms, so a bad match is dismissible at a glance. Never blocks: any error exits 0 silently.

How selection works

Both scorers rank every document. Their rankings are combined by reciprocal-rank fusion — each scorer contributes 1 / (60 + rank) — and the top two documents are returned. A document ranked moderately by both scorers therefore beats one ranked first by only one, which is the whole reason to run a hybrid.

lex_threshold and sem_threshold are admission floors ("is this document a candidate at all"), not the selection rule. They are what makes abstention possible: a prompt neither scorer admits routes to nothing.

Absolute score cutoffs are deliberately not the selection rule. TF-IDF cosine and embedding cosine are not on a common scale, and the right cutoff for either moves with the corpus. Through 0.1.2 oiax selected on absolute cutoffs with a semantic threshold of 0.55; on the reference corpus, correct semantic matches score 0.40–0.48, so the semantic half never fired and recall sat at 0.185. Rank fusion is scale-free.

Defaults are calibrated, not chosen: src/oiax/eval/corpora/README.md records the sweep, the operating point, and what it was picked over.

Evaluation harness

Measure routing quality against labelled ground truth:

python -m oiax.eval.route_eval score ./policies/ < labelled.jsonl   # shipped config
python -m oiax.eval.route_eval sweep ./policies/ < labelled.jsonl   # the full grid

The labelled file is JSONL — one JSON object per line:

{"prompt": "How do I deploy to production?", "expected": ["deploy-policy"]}
{"prompt": "What's for lunch?", "expected": []}

Reported: recall@2, precision, F1, top-1 accuracy, and the false-alarm rate over negative prompts ("expected": []). Read precision against the two-hit cap — with one expected label it cannot exceed 0.5 for that prompt.

Two corpora ship at oiax/eval/corpora/: a 15-document reference corpus with 52 labelled prompts (the calibration set — recall@2 0.648, top-1 0.673, zero false alarms), and a 5-document synthetic smoke corpus that is structurally useful and cannot calibrate anything. Judge labels are evidence, not proof — hand-check a slice before treating any rate as authoritative.

API

oiax.router

Callable	Signature	Returns
route	`route(prompt: str, index: Index	None = None) -> list[RouteHit]`
build_index	build_index(corpus, *, expansions, lex_threshold, sem_threshold, rrf_k, top_k) -> Index	Built index
semantic_ready	semantic_ready() -> bool	False when the embedding model failed to load and routing is lexical-only — surface it, do not swallow it

RouteHit

@dataclass(frozen=True)
class RouteHit:
    name: str       # document name (surface name only, never body text)
    score: float    # best RAW scorer score, [0, 1] — hits are ORDERED by fusion, not by this
    why: list[str]  # matched terms, and/or "semantic match"

oiax.corpus

Class	Purpose
Document(name, trigger_line, body)	One document in the routing corpus
Corpus (Protocol)	Any object with .documents() -> Iterator[Document]
PolicyDirCorpus(path)	Reads *.md files with **Agent-trigger:** headers

oiax.adapters

Module	Purpose
claude_code.py	UserPromptSubmit hook adapter
stdout.py	Debug adapter — prints hits as text

When you need oiax

You need oiax when your rule corpus is too large to inject into every context (context-window pressure, attention dilution) and too important to leave to the agent's judgment (silent policy violations).

You do not need oiax when your corpus fits in a single CLAUDE.md — static injection is free and optimal for that case.

Development

git clone https://github.com/nousergon/oiax.git
cd oiax
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

pytest                          # test suite
ruff check src/ tests/          # lint
mypy src/oiax                   # type check

All three run in CI on Python 3.11, 3.12 and 3.13 and are required to merge. See CONTRIBUTING.md.

License

AGPL-3.0 — see LICENSE.