yzhao062 / pyod / 26122516955

93%

master: 93%

push

github

v3.5.3: KB-tools API for agent-driven and LLM-driven routing

Surface 1 (agent tools):
- ADEngine.get_kb_for_routing(profile, top_k=3, constraints=None) returns a
  structured KB snapshot (every shipped detector with strengths, weaknesses,
  best_for, avoid_when, complexity, benchmark_rank, modality_match) filtered
  by exclude_detectors + data_type_strict and sorted by modality-specific
  benchmark rank keys (TSB_AD_overall for time-series, BOND_deep for graph,
  NLP_ADBench_overall for text, MVTec_overall for image, ADBench_overall
  fallback). Each entry carries resolved_rank + resolved_rank_key for
  downstream tools.
- ADEngine.make_plan(detector_choices, justifications=None, params=None)
  validates names against the KB (case-sensitive, must be shipped) and
  returns a closed-schema DetectionPlan consumable by build_detector / run.

Surface 2 (programmatic API):
- ADEngine.plan_detection(profile, priority='balanced', constraints=None, *,
  top_k=3, llm_client=None, llm_strict=None) accepts a (prompt: str) -> str
  callable. Engine builds the routing prompt internally, invokes the
  callable, parses the response, enforces the constrained KB context
  post-parse (LLM cannot bypass exclude_detectors / data_type_strict), and
  returns a DetectionPlan with note='llm-driven via plan_detection(...)' +
  evidence=['llm_routing']. On LLM call or parse failure, falls back to
  rule routing with a RuntimeWarning. llm_strict=True per-call overrides
  the PYOD3_LLM_STRICT env var; precedence is explicit kwarg > env > default.
- pyod/utils/_llm.py: new module with LLMCallable Protocol, RoutingParseError,
  build_routing_prompt(kb_context, top_k), parse_routing_response(response,
  kb, top_k). Parser tolerates surrounding prose, markdown fences, BOM/CRLF,
  skips unknown / non-shipped detectors with a logged warning, dedupes,
  truncates to top_k.

top_k generalization: plan_detection's previous valid[1:3] hard cap is now
valid[1:top_k]. Default 3 preserves v3... (continued)

483 of 502 new or added lines in 3 files covered. (96.22%)

18589 of 19802 relevant lines covered (93.87%)

10.3 hits per line