feat(LongTermMemory): retain fact provenance and review model-selected evidence

This commit is contained in:
huanghuoguoguo
2026-11-02 20:45:14 +08:00
parent 7d1f217350
commit 7cd90087b1
11 changed files with 784 additions and 9 deletions
@@ -23,6 +23,7 @@ from langbot_plugin.api.entities.builtin.rag.models import (
from langbot_plugin.api.entities.builtin.rag.enums import DocumentStatus
from store.recall import matches_filters, select_memories
from store.serialization import serialized_write
from store.episode_provenance import normalize_provenance
logger = logging.getLogger(__name__)
@@ -347,6 +348,15 @@ class LongTermMemoryEngine(KnowledgeEngine):
chunks_created=0,
)
try:
if any(not isinstance(mem, dict) for mem in memories):
raise ValueError("each imported memory must be an object")
provenances = [normalize_provenance(mem) for mem in memories]
except ValueError as exc:
return IngestionResult(
document_id=doc_id, status=DocumentStatus.FAILED, error_message=str(exc),
)
# batch embed and upsert
total_stored = 0
batch_texts: list[str] = []
@@ -354,7 +364,7 @@ class LongTermMemoryEngine(KnowledgeEngine):
batch_metas: list[dict[str, Any]] = []
imported_by_user: dict[str, int] = {}
for mem in memories:
for mem, provenance in zip(memories, provenances):
content = mem.get("content", "")
if not content:
continue
@@ -382,10 +392,13 @@ class LongTermMemoryEngine(KnowledgeEngine):
user_key_text = str(user_key or "imported")
imported_by_user[user_key_text] = imported_by_user.get(user_key_text, 0) + 1
if not provenance["source_ref"]:
provenance["source_ref"] = f"document:{doc_id}"
mid = uuid_mod.uuid4().hex[:12]
batch_texts.append(content)
batch_ids.append(mid)
batch_metas.append({
**provenance,
"episode_id": mid,
"content": content,
"tags": ",".join(tags) if isinstance(tags, list) else str(tags),
@@ -164,6 +164,17 @@ spec:
required: false
default: false
- name: candidate_model_uuid
label:
en_US: Candidate Selection Model
zh_Hans: 候选筛选模型
description:
en_US: Optional model for selecting original evidence spans. Leave empty for rules. Model-selected candidates require manual review and are never auto-applied.
zh_Hans: 可选的原文证据筛选模型。留空使用规则。模型筛选结果需要人工审核,不会自动应用。
type: llm-model-selector
required: false
default: ''
- name: candidate_auto_apply
label:
en_US: Auto Apply Memory Candidates