Retire the request-time hint API. PR-B5 splits TOIN into two phases:
1. Observation: TOIN keeps recording compressions/retrievals at runtime,
but `get_recommendation()` is deprecated and now returns None.
2. Publish-then-load: the new `headroom.cli.toin_publish` CLI walks the
on-disk store and emits `recommendations.toml`. The Rust proxy reads
that file once at startup via `transforms::recommendations` and
exposes `get(auth_mode, model, structure_hash) -> Option<&Rec>`.
PR-F3 will wire the loader into the live-zone dispatcher.
Per-tenant aggregation: `_patterns` is now keyed by
`(auth_mode, model_family, sig_hash)` so PAYG/OAuth/subscription tenants
no longer share buckets. Callers that don't supply auth/model land in the
`("unknown", "unknown", sig_hash)` slot. Added `_make_pattern_key` helper
+ updated tests that previously indexed by raw `structure_hash`.
AuthMode is canonical in `transforms::live_zone`; `transforms::recommendations`
re-exports it (no duplicate enum). Live-zone enum gained `Unknown`,
`as_str()`, and `Hash` derive to serve recommendations callers without a
second source of truth.
Why: per-request hint calls coupled output to mutable TOIN state, breaking
prompt-cache stability across runs (P2-27, P5-56). Pulling advice into a
startup-published TOML keeps per-request output deterministic and lets the
deploy pipeline gate publication independently of proxy uptime.
Per-PR-B5 plan: REALIGNMENT/04-phase-B-live-zone.md.
- Add headroom.evals module with 12+ dataset loaders (HotpotQA, SQuAD,
Natural Questions, TriviaQA, MS MARCO, LongBench, NarrativeQA, BFCL,
ToolBench, CodeSearchNet, HumanEval, built-in tool outputs)
- Add before/after evaluation runner that compares LLM responses with
original vs compressed context
- Add metrics: F1 score, semantic similarity, exact match, ground truth
- Add CLI: python -m headroom.evals quick|benchmark|list|report
- Add [evals] extra to pyproject.toml for pip install headroom-ai[evals]
Fix ContentRouter to use LLMLingua for plain text compression:
- Route TEXT strategy through LLMLingua instead of heuristic TextCompressor
- Adjust LLMLingua compression rates for better accuracy (0.5 vs 0.25)
- HotpotQA now achieves 95% accuracy with 44% compression
Update documentation with evaluation framework section
Fix test isolation in test_toin.py (TOIN singleton persistence)
Enable TOIN (Tool Output Intelligence Network) to work end-to-end
through the proxy by fixing content routing and adding persistence.
Changes:
- Fix ContentRouter to route json_array hint to SmartCrusher
(prevents JSON arrays with text from being misclassified as mixed)
- Add default TOIN storage path (~/.headroom/toin.json)
- Support HEADROOM_TOIN_PATH env var for custom storage location
- Add TOIN API endpoints: /v1/toin/stats, /v1/toin/patterns,
/v1/toin/pattern/{hash_prefix}
- Lower min_samples threshold from 10 to 3 for faster learning
- Add progressive confidence based on sample count
- Add debug logging for TOIN hint application
- Add comprehensive integration tests (no mocks)
