$j)UdZddlZddlmZmZddlmZmZejZdZ e e d<dZ e e d<d Z e e d <e e e d fd eed ede de de deddfdZGddZdS)z4 Cross-encoder neural reranking for search results. N)datetimetimezone)MergedCandidate ScoredResultg?_RECENCY_ALPHA_TEMPORAL_ALPHA皙?_PROOF_COUNT_ALPHAFscored_resultsnow recency_alphatemporal_alphaproof_count_alphais_passthrough_rerankerreturnc |j|t}|r\|rZt|}t |dd}t d|dz }t |D]\} } dd| z|z z | _|D]F} d | _| j j rq| j j } | j| t} || z d z } t d tdd| d z z | _| j j | j j nd | _| j j} | =| dkr7tdt d d t!j| dz z}nd }d | _d|| jd z zz}d|| jd z zz}d||d z zz}| j|z|z|z| _| j| _HdS)uApply combined scoring to a list of ScoredResults in-place. Uses the cross-encoder score as the primary relevance signal, with recency, temporal proximity, and proof count applied as multiplicative boosts. This ensures the influence of these secondary signals is always proportional to the base relevance score, regardless of the cross-encoder model's score calibration. Formula:: recency_boost = 1 + recency_alpha * (recency - 0.5) # in [1-α/2, 1+α/2] temporal_boost = 1 + temporal_alpha * (temporal - 0.5) # in [1-α/2, 1+α/2] proof_count_boost = 1 + proof_count_alpha * (proof_norm - 0.5) # in [1-α/2, 1+α/2] combined_score = CE_normalized * recency_boost * temporal_boost * proof_count_boost proof_norm maps proof_count using a smooth logarithmic curve centered at 0.5, clamped to [0, 1]: proof_count=1 → 0.5 + 0 = 0.5 (neutral multiplier) proof_count=150 → clamped to 1.0 (max +5% boost) Temporal proximity is treated as neutral (0.5) when not set by temporal retrieval, so temporal_boost collapses to 1.0 for non-temporal queries. Proof count is treated as neutral (0.5) when not available (non-observation facts), so proof_count_boost collapses to 1.0 for world/experience/opinion facts. Args: scored_results: Results from the cross-encoder reranker. Mutated in place. now: Current UTC datetime for recency calculation. recency_alpha: Max relative recency adjustment (default 0.2 → ±10%). temporal_alpha: Max relative temporal adjustment (default 0.2 → ±10%). proof_count_alpha: Max relative proof count adjustment (default 0.1 → ±5%). N)tzinfocBtt|ddddS)N candidate rrf_score)getattr)ss m/home/rurouni/.hermes/hermes-agent/venv/lib/python3.11/site-packages/hindsight_api/engine/search/reranking.pyz(apply_combined_scoring..[s''![$"?"?cRRTkeyreverserg?g?g?iQr imrg$@)rreplaceUTClensortedmax enumeratecross_encoder_score_normalizedrecency retrievaloccurred_start total_secondsmintemporal_proximitytemporal proof_countmathlogrrf_normalizedcombined_scoreweight)r r rrrrn sorted_by_rrfdenomnew_ranksroccurreddays_agor/ proof_norm recency_boosttemporal_boostproof_count_boosts rapply_combined_scoringr@s#R zkkk%%2 O> O    RR   Aq1u %m44 O OLHb14sX~7M0NB - -&& < & D|2H&#++3+77h5577%?HS#c3(S.+A"B"BCCBJ:<9X9dbl55jm l.  "{a'7'7S#c3$(;2G2G$2N+O"P"PQQJJJ mrzC/?@@ ~s1BCC"3zC7G"HH= MP^^arr% =&&rcLeZdZdZd dZdZdedeedee fdZ dS) CrossEncoderRerankerz Neural reranking using a cross-encoder model. Configured via environment variables (see cross_encoder.py). Default local model is cross-encoder/ms-marco-MiniLM-L-6-v2. NcF|ddlm}|}||_d|_dS)z Initialize cross-encoder reranker. Args: cross_encoder: CrossEncoderModel instance. If None, creates one from environment variables (defaults to local provider) Nr)create_cross_encoder_from_envF)"hindsight_api.engine.cross_encoderrD cross_encoder _initialized)selfrFrDs r__init__zCrossEncoderReranker.__init__s@  X X X X X X99;;M*!rcK|jrdSddl|jjdkr5}|dfdd{Vnd{Vd|_dS)zHEnsure the cross-encoder model is initialized (for lazy initialization).NrlocalcRSN)run initialize)asynciorFsrrz9CrossEncoderReranker.ensure_initialized..sW[[AYAYA[A[5\5\rT)rGrPrF provider_nameget_event_looprun_in_executorrO)rHlooprPrFs @@rensure_initializedz'CrossEncoderReranker.ensure_initializeds    F*  &' 1 1))++D&&t-\-\-\-\-\]] ] ] ] ] ] ] ] ]**,, , , , , , , , rquery candidatesrcK|sgSg}|D]}}|j}|j}|jr |jd|}|jr<|j}|d}|d} d| d|d|}|||g~|j|d{V} ddl} ddl fd fd | D} g} t|| | D]p\}}}t|}t|}| j |rd }| j |rd }t|||| }| |q| d d| S)z Rerank candidates using cross-encoder scores. Args: query: Search query candidates: Merged candidates from RRF Returns: List of ScoredResult objects sorted by cross-encoder score z: z%Y-%m-%dz %B %d, %Yz[Date: z (z)] Nrc<dd| zz S)Nr)exp)xnps rsigmoidz,CrossEncoderReranker.rerank..sigmoidsBFFA2JJ' 'rc&g|] }|Sr_).0scorer]s r z/CrossEncoderReranker.rerank..s!@@@WWU^^@@@rr)rcross_encoder_scorer'r4c|jSrM)r4)r[s rrz-CrossEncoderReranker.rerank..s!(rTr)r)textcontextr*strftimeappendrFpredictr0numpyzipfloatisnanrsort)rHrVrWpairsrr)doc_textr*date_iso date_readablescoresr0normalized_scoresr raw_score norm_scorerawnorm scored_resultr\r]s @@rrerankzCrossEncoderReranker.reranks$ I# , ,I!+I!~H  >'/==8==' N!*!9*22:>>!/ 7 7 D D N]MMhMM8MM LL%* + + + +)11%88888888   ( ( ( ( (A@@@@@@03JHY0Z0Z 1 1 ,Iy* ""C$$Dtz# tz$ (#$'/3 M  ! !- 0 0 0 0  2 2DAAArrM) __name__ __module__ __qualname____doc__rIrUstrlistrrrzr_rrrBrBs~ " " " "!!! M#M43HMTR^M_MMMMMMrrB)r~r0rrtypesrrutcr"rrl__annotations__r r rboolr@rBr_rrrs@ ''''''''00000000l E *+1$) n&n&&n& n&n& n&  n& " n& n&n&n&n&bttttttttttr