|
| 1 | +from __future__ import annotations |
| 2 | + |
| 3 | +import logging |
| 4 | + |
| 5 | +from dataclasses import dataclass |
| 6 | +from typing import TYPE_CHECKING, Any |
| 7 | + |
| 8 | +from memos.dream.contextualization import CONTEXT_MEMORY_TYPE |
| 9 | + |
| 10 | + |
| 11 | +if TYPE_CHECKING: |
| 12 | + from memos.api.product_models import APISearchRequest |
| 13 | + |
| 14 | + |
| 15 | +logger = logging.getLogger(__name__) |
| 16 | + |
| 17 | +_DEFAULT_CONTEXT_RECALL_TOP_K = 2 |
| 18 | +_CONTEXT_RETURN_FIELDS = [ |
| 19 | + "memory", |
| 20 | + "key", |
| 21 | + "created_at", |
| 22 | + "updated_at", |
| 23 | + "source", |
| 24 | + "internal_info", |
| 25 | +] |
| 26 | + |
| 27 | + |
| 28 | +@dataclass |
| 29 | +class DreamContextSearchExtension: |
| 30 | + """Dream-owned search extension for recalling Context nodes. |
| 31 | +
|
| 32 | + The core SearchHandler only exposes a generic plugin hook. This extension |
| 33 | + owns Dream-specific retrieval details such as the Context memory type, |
| 34 | + graph scope, metadata formatting, and fallback behavior. |
| 35 | + """ |
| 36 | + |
| 37 | + top_k: int = _DEFAULT_CONTEXT_RECALL_TOP_K |
| 38 | + |
| 39 | + def merge_context_recall( |
| 40 | + self, |
| 41 | + *, |
| 42 | + handler, |
| 43 | + search_req: APISearchRequest, |
| 44 | + results: dict[str, Any], |
| 45 | + ) -> dict[str, Any]: |
| 46 | + top_k = max(0, int(self.top_k or 0)) |
| 47 | + if top_k <= 0: |
| 48 | + return results |
| 49 | + |
| 50 | + context_buckets = self._recall_context_buckets( |
| 51 | + handler=handler, |
| 52 | + search_req=search_req, |
| 53 | + top_k=top_k, |
| 54 | + ) |
| 55 | + if context_buckets: |
| 56 | + results.setdefault("text_mem", []).extend(context_buckets) |
| 57 | + return results |
| 58 | + |
| 59 | + def _recall_context_buckets( |
| 60 | + self, *, handler, search_req: APISearchRequest, top_k: int |
| 61 | + ) -> list[dict[str, Any]]: |
| 62 | + graph_db = getattr(handler, "graph_db", None) or getattr( |
| 63 | + handler.searcher, "graph_store", None |
| 64 | + ) |
| 65 | + embedder = getattr(handler, "embedder", None) or getattr(handler.searcher, "embedder", None) |
| 66 | + if graph_db is None or embedder is None: |
| 67 | + logger.info("[Dream Search] Context recall skipped: graph_db or embedder unavailable.") |
| 68 | + return [] |
| 69 | + |
| 70 | + try: |
| 71 | + query_embedding = embedder.embed([search_req.query])[0] |
| 72 | + except Exception: |
| 73 | + logger.warning("[Dream Search] Context recall embedding failed.", exc_info=True) |
| 74 | + return [] |
| 75 | + |
| 76 | + buckets: list[dict[str, Any]] = [] |
| 77 | + for cube_id in _resolve_cube_ids(search_req): |
| 78 | + try: |
| 79 | + hits = graph_db.search_by_embedding( |
| 80 | + query_embedding, |
| 81 | + top_k=top_k, |
| 82 | + scope=CONTEXT_MEMORY_TYPE, |
| 83 | + status="activated", |
| 84 | + user_name=cube_id, |
| 85 | + return_fields=_CONTEXT_RETURN_FIELDS, |
| 86 | + ) |
| 87 | + except Exception: |
| 88 | + logger.warning( |
| 89 | + "[Dream Search] Context recall search failed for cube=%s.", |
| 90 | + cube_id, |
| 91 | + exc_info=True, |
| 92 | + ) |
| 93 | + continue |
| 94 | + |
| 95 | + memories = [_format_context_hit(hit) for hit in hits or [] if hit.get("memory")] |
| 96 | + if not memories: |
| 97 | + continue |
| 98 | + buckets.append( |
| 99 | + { |
| 100 | + "cube_id": cube_id, |
| 101 | + "memories": memories, |
| 102 | + "total_nodes": len(memories), |
| 103 | + } |
| 104 | + ) |
| 105 | + return buckets |
| 106 | + |
| 107 | + |
| 108 | +def _resolve_cube_ids(search_req: APISearchRequest) -> list[str]: |
| 109 | + if search_req.readable_cube_ids: |
| 110 | + return list(dict.fromkeys(search_req.readable_cube_ids)) |
| 111 | + return [search_req.user_id] |
| 112 | + |
| 113 | + |
| 114 | +def _format_context_hit(hit: dict[str, Any]) -> dict[str, Any]: |
| 115 | + context_id = str(hit.get("id", "")) |
| 116 | + score = float(hit.get("score", 0.0) or 0.0) |
| 117 | + metadata = { |
| 118 | + "id": context_id, |
| 119 | + "memory": hit.get("memory", ""), |
| 120 | + "memory_type": CONTEXT_MEMORY_TYPE, |
| 121 | + "source": hit.get("source") or "dream", |
| 122 | + "key": hit.get("key", ""), |
| 123 | + "relativity": score, |
| 124 | + "score": score, |
| 125 | + "embedding": [], |
| 126 | + "sources": [], |
| 127 | + "usage": [], |
| 128 | + "ref_id": f"[{context_id.split('-')[0]}]" if context_id else "[context]", |
| 129 | + } |
| 130 | + for field in ("created_at", "updated_at", "internal_info"): |
| 131 | + if hit.get(field) is not None: |
| 132 | + metadata[field] = hit[field] |
| 133 | + |
| 134 | + return { |
| 135 | + "id": context_id, |
| 136 | + "memory": hit.get("memory", ""), |
| 137 | + "metadata": metadata, |
| 138 | + "ref_id": metadata["ref_id"], |
| 139 | + } |
0 commit comments