src.dackar.RCA.ner.ner_adapter

Attributes

_SUPPORTED_MODES

Functions

_build_token_index(gaz[, min_tok_len])

Invert gazetteer label_terms into token -> {(label, term)} for AnchoredNPGenerator.

build_ner_pipeline(schema_json_path, gazetteer_xl, ...)

Build and return a HybridNERPipeline.

_uniq(seq)

_route_entity(schema, labels, groups)

Return one canonical NERSeed bucket:

ner_seed_provider_from_pipeline(pipeline, NERSeed[, ...])

build_ner_provider(schema_json_path, gazetteer_xl, ...)

Convenience factory: build pipeline + SpacyAnnotator and return a ready-to-use

Module Contents

src.dackar.RCA.ner.ner_adapter._SUPPORTED_MODES[source]
src.dackar.RCA.ner.ner_adapter._build_token_index(gaz, min_tok_len=3)[source]

Invert gazetteer label_terms into token -> {(label, term)} for AnchoredNPGenerator.

Parameters:
Return type:

Dict[str, set]

src.dackar.RCA.ner.ner_adapter.build_ner_pipeline(schema_json_path, gazetteer_xl, label_json, llm_cfg=None, generator_mode='gazetteer_only', np_score_threshold=0.65)[source]

Build and return a HybridNERPipeline.

generator_mode options

"gazetteer_only" (default)

Only exact-phrase gazetteer matching. Highest precision, lowest recall. DescriptionEmbedGenerator classifies gazetteer hits; OOV spans are invisible.

"anchored_np" (recommended upgrade)

Gazetteer + AnchoredNPGenerator. Noun-phrase spans are proposed only when they contain at least one gazetteer token, so OOV compound phrases like “fretting corrosion” or “oxide layer buildup” become candidates while purely generic NPs (“work order”, “inspection notes”) are suppressed. DescriptionEmbedGenerator then classifies all candidates with its Rule 1/2 gates; np_score_threshold (default 0.65) is applied instead of the default 0.55 to keep precision acceptable.

"full_np"

Gazetteer + plain NounPhraseGenerator. Every spaCy noun chunk becomes a candidate regardless of gazetteer overlap. Highest recall, more noise. Use for exploratory indexing or when vocabulary coverage is thin.

param np_score_threshold:

Minimum cosine similarity accepted by DescriptionEmbedGenerator when NP generation is active ("anchored_np" or "full_np"). Ignored for "gazetteer_only" / "default".

Parameters:

  • schema_json_path (str)

  • gazetteer_xl (str)

  • label_json (str)

  • llm_cfg (dict)

  • generator_mode (str)

  • np_score_threshold (float)

Return type:

src.dackar.RCA.ner.hybrid_ner.pipeline.HybridNERPipeline

src.dackar.RCA.ner.ner_adapter._uniq(seq)[source]
Parameters:

seq (Iterable[str])

Return type:

list[str]

src.dackar.RCA.ner.ner_adapter._route_entity(schema, labels, groups)[source]
Return one canonical NERSeed bucket:

systems, components, mechanisms, outcomes, maintenance_actions, surveillance_actions, tools, properties

Parameters:

  • schema (Any)

  • labels (list[str])

  • groups (list[str])

Return type:

Optional[str]

src.dackar.RCA.ner.ner_adapter.ner_seed_provider_from_pipeline(pipeline, NERSeed, annotator=None)[source]
Parameters:
Return type:

Callable[[Dict[str, Any]], Any]

src.dackar.RCA.ner.ner_adapter.build_ner_provider(schema_json_path, gazetteer_xl, label_json, NERSeed, llm_cfg=None, generator_mode='anchored_np', np_score_threshold=0.65, spacy_model='en_core_web_sm')[source]

Convenience factory: build pipeline + SpacyAnnotator and return a ready-to-use NER seed provider callable.

Intended for use with augment_chunks_with_structured_summaries():

provider = build_ner_provider(schema, gaz, label, NERSeed)
augment_chunks_with_structured_summaries(chunks_path, ner_seed_provider=provider, ...)

Parameters mirror build_ner_pipeline(). spacy_model selects the base spaCy model for the SpacyAnnotator (Tier 1 measurement/temporal/location).

Parameters:

  • schema_json_path (str)

  • gazetteer_xl (str)

  • label_json (str)

  • llm_cfg (dict)

  • generator_mode (str)

  • np_score_threshold (float)

  • spacy_model (str)

Return type:

Callable[[Dict[str, Any]], Any]