src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher

PatternSearcher — coarse-to-fine similarity retrieval pipeline.

Executes three-step retrieval against a pre-built IncidentIndex:

1. Inverted-index lookup — O(|query_event_set|) candidate set

2. Jaccard pre-filter — discard clearly dissimilar episodes

3. Full metric scoring — NLCS + EMD + combined weighted score

All three metric scores are always computed and returned so that callers can switch weight profiles or analyse individual signals without re-running.

Attributes

_log
_INDEX_STATUS_INDEXED
_INDEX_STATUS_NO_DATA
_INDEX_STATUS_STALE

Classes

PatternSearcher

Retrieves the top-k most similar historical episodes for a query incident.

Functions

_compute_index_status(index, staleness_window_days)	Return the index_status string for a given index and staleness window.
_make_no_data_sentinel(asset_id[, index_status])	Return a sentinel HistoricalSignalEpisode for the no-data case.

Module Contents

src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher._log

src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher._INDEX_STATUS_INDEXED = 'indexed'

src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher._INDEX_STATUS_NO_DATA = 'no_episodes_indexed'

src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher._INDEX_STATUS_STALE = 'stale'

class src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher.PatternSearcher(index, config)

Retrieves the top-k most similar historical episodes for a query incident.

Pipeline per search() call:

1. Inverted-index lookup: episode_ids sharing ≥ 1 event type with query.

2. Jaccard pre-filter: discard candidates below config.min_jaccard.

3. NLCS computation on survivors.

4. EMD computation on survivors.

5. Combined score: weighted sum using the resolved weight profile.

6. Rank descending by combined_score; return the top-k HistoricalSignalEpisodes.

The coarse-to-fine design avoids computing NLCS and EMD on clearly dissimilar episodes (those failing the Jaccard gate).

Parameters:

• index (src.dackar.RCA.log_pattern_recognition.rca_pattern_search.indexer.IncidentIndex)

• config (src.dackar.RCA.log_pattern_recognition.rca_pattern_search.config.SearchConfig)

index

config

search(query, weight_profile=None, staleness_window_days=None)

Retrieves top-k most similar historical episodes for a query fingerprint.

Returns list[HistoricalSignalEpisode] with index_status populated on every result (§4.11):

• “indexed” — normal result from a current, populated index

• “no_episodes_indexed” — index is empty; returns a single sentinel episode

• “stale” — index is older than staleness_window_days; results

  returned but flagged; link_confidence capped downstream

Parameters:

• query (src.dackar.RCA.log_pattern_recognition.rca_pattern_search.models.IncidentFingerprint) – IncidentFingerprint for the query incident.

• weight_profile (Optional[str]) – Weight profile override; None → config.weight_profile.

• staleness_window_days (Optional[int]) – If set and index.build_timestamp is known, episodes from an index older than this are marked “stale”. None disables the staleness check. By design the window lives on PatternSearchConfig.index_staleness_window_days; the orchestrator reads it there and passes it in here (PatternSearcher itself only carries a SearchConfig).

Return type:

list[src.dackar.RCA.log_pattern_recognition.rca_pattern_search.models.HistoricalSignalEpisode]

Notes

• All three metric scores (jaccard, nlcs, emd) are individually visible on every returned HistoricalSignalEpisode (§5).

• matched_events, query_only_events, episode_only_events are derived from event_set comparison.

_resolve_weights(weight_profile)

Returns (alpha, beta_w, gamma) for the given weight profile name.

Delegates to SearchConfig.resolve_weights() which owns the profile registry and “custom” fallback logic.

Raises:

ValueError for unrecognised profile names. –

Parameters:

weight_profile (str)

Return type:

tuple[float, float, float]

src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher._compute_index_status(index, staleness_window_days)

Return the index_status string for a given index and staleness window.

Parameters:

• index (src.dackar.RCA.log_pattern_recognition.rca_pattern_search.indexer.IncidentIndex)

• staleness_window_days (Optional[int])

Return type:

str

src.dackar.RCA.log_pattern_recognition.rca_pattern_search.searcher._make_no_data_sentinel(asset_id, index_status=_INDEX_STATUS_NO_DATA)

Return a sentinel HistoricalSignalEpisode for the no-data case.

Callers must check index_status before attempting cross-pattern linkage. A sentinel has episode_id == “” and similarity_to_current == 0.0.

Parameters:

• asset_id (str)

• index_status (str)

Return type:

src.dackar.RCA.log_pattern_recognition.rca_pattern_search.models.HistoricalSignalEpisode