iwdZddlZddlZddlmZmZddlmZmZddlm Z eje Z dZ GddZd ed efd Zy) u core/workers/memory_capture_worker.py Story 5.08: MemoryCaptureWorker — Utterance-to-KG Entity AIVA RLM Nexus PRD v2 — Track A Immediate utterance capture: does NOT wait for call.hangup. Fires when AIVA detects a CAPTURE_MEMORY intent mid-call. Use for: Kinan statements of fact, new directives, critical decisions. Design notes: - Writes to the `aiva_conversations` Qdrant collection immediately. - Completes in < 2s (no Gemini call, no Redis dependency). - Falls back to a deterministic SHA-256-based embedding if no embedding client is injected (matches the pattern used in PostCallEnricher). - All external I/O (Qdrant, embedding) is injected — zero hardwired side effects, fully testable without real services. N)datetimetimezone)AnyOptional)uuid4aiva_conversationscReZdZdZ d dededdfdZdefdZdefdZd e de fd Z y) MemoryCaptureWorkera Immediate utterance capture: does NOT wait for call.hangup. Fires when AIVA detects a CAPTURE_MEMORY intent mid-call. Use for: Kinan statements of fact, new directives, critical decisions. Usage: worker = MemoryCaptureWorker(qdrant_client=qdrant, embedding_client=embedder) result = await worker.execute(intent_signal) # {"entity_id": "", "status": "captured"} N qdrant_clientembedding_clientreturnc ||_||_y)a Args: qdrant_client: A Qdrant client with upsert(collection_name, points) method. If None, the upsert is skipped (returns status="captured" anyway so callers always receive a well-formed response, but a warning is logged). embedding_client: An object with an embed(text: str) -> list[float] method. If None, a deterministic SHA-256-based 768-dim fallback is used. N)_qdrant _embedding)selfr r s ;/mnt/e/genesis-system/core/workers/memory_capture_worker.py__init__zMemoryCaptureWorker.__init__+s% *cKtt}|j|}|j|jd{}|j T ddlm}|j jt||||gtjd||jntj!d ||d d S7#t$rJ|j jt|||dgtjd||jYYt$r!}tjd||Yd}~}d}~wwxYww) a  Capture an important utterance as a KG entity in Qdrant immediately. Steps: 1. Generate a UUID for the new Qdrant point. 2. Build the entity payload from intent fields. 3. Embed intent.utterance into a 768-dim vector. 4. Upsert the point to the aiva_conversations Qdrant collection. 5. Return {"entity_id": "", "status": "captured"}. Args: intent: An IntentSignal with at minimum: - intent.utterance (str) - intent.session_id (str) - intent.intent_type (IntentType) - intent.extracted_entities (dict) Returns: dict with keys "entity_id" (str UUID) and "status" ("captured"). Nr) PointStruct)idvectorpayload)collection_namepointsz>MemoryCaptureWorker.execute: upserted entity %s for session %szKMemoryCaptureWorker.execute: upserted entity %s (duck-typed) for session %suQMemoryCaptureWorker.execute: Qdrant upsert failed for entity %s: %s — non-fataluRMemoryCaptureWorker.execute: no Qdrant client injected — entity %s not persistedcaptured) entity_idstatus)strr_build_entity_payload _embed_text utterancerqdrant_client.modelsrupsertQDRANT_COLLECTIONloggerinfo session_id ImportError Exceptionerrorwarning)rintentrrrrexcs rexecutezMemoryCaptureWorker.execute<s6*L ,,V4''(8(899 << # < ##$5#yQ$  T%%. NNd  '*==Q:"  ##$5#,7ST$ a%%   g sJAECEAC)EAEEE!D=8E=EEc Ht|jdr|jjnt|j}|j|j t jtjj|t|dit|dddS)aK Build the payload dict stored as the Qdrant point's payload. Required fields per acceptance criteria: - utterance (str): the raw caller utterance - session_id (str): the AIVA/Telnyx call session identifier - captured_at (str): ISO 8601 UTC timestamp of capture - intent_type (str): the string value of the IntentType enum Additional fields (enrichment): - extracted_entities (dict): any entities the IntentClassifier extracted - confidence (float): classifier confidence score valueextracted_entities confidence)r"r( captured_at intent_typer2r3) hasattrr6r1rr"r(rnowrutc isoformatgetattr)rr-intent_type_strs rr z)MemoryCaptureWorker._build_entity_payload~s"v))73    $ $V''(  )) ++#<< 5??A*")&2F"K!&,<   rtextcnK|j|jj|St|Sw)a Generate a 768-dim embedding vector for the given text. If an embedding_client was injected, delegates to ``embedding_client.embed(text)`` which must return a list[float]. Otherwise falls back to the same deterministic SHA-256-based embedding used in PostCallEnricher (for testing + zero-dependency operation). Args: text: The utterance string to embed. Returns: list of 768 floats. )rembed _sha256_embed)rr=s rr!zMemoryCaptureWorker._embed_texts2" ?? &??((. .T""s35)NN) __name__ __module__ __qualname____doc__rrdictr/r rlistr!rrr r s^ " $+++  +"@>t@>D t <#c#d#rr r=r c tj|jj}|Dcgc]}t |ddz }}|dddgt ddt |z zzScc}w)u Deterministic 768-dim embedding derived from SHA-256. This is a zero-dependency fallback suitable for tests and environments where no real embedding model is available. The vector is NOT semantically meaningful — swap ``MemoryCaptureWorker._embed_text`` body with a real model call for production use. Per the PRD specification pattern: digest = hashlib.sha256(text.encode()).hexdigest() return [int(c, 16) / 15.0 for c in digest][:768] + [0.0] * max(0, 768 - 64) g.@Nir4r)hashlibsha256encode hexdigestintmaxlen)r=digestc hex_valuess rr@r@sm^^DKKM * 4 4 6F-34#a*t#4J4 ds sec!S3z?-B&CC CC5sA0)rDrJloggingrrtypingrruuidr getLoggerrAr&r%r rrFr@rGrrrXsV&'    8 $(Q#Q#hDDDr