20260625:很多新内容
This commit is contained in:
36
concepts/verbatim-pre-recall.md
Normal file
36
concepts/verbatim-pre-recall.md
Normal file
@@ -0,0 +1,36 @@
|
||||
---
|
||||
title: "Verbatim Pre-Recall"
|
||||
created: 2026-06-24
|
||||
updated: 2026-06-24
|
||||
type: concept
|
||||
tags: ["information-retrieval", "agent-design", "bm25", "query-processing"]
|
||||
sources:
|
||||
- "[[atlas-agent-memory-architecture-2026]]"
|
||||
---
|
||||
|
||||
# Verbatim Pre-Recall
|
||||
|
||||
Verbatim Pre-Recall 是 Atlas 记忆系统的关键设计:在 LLM 看到用户消息之前,先用用户原话(一字不改)跑一次检索。这是保障 BM25 词法腿价值的关键。
|
||||
|
||||
## 问题
|
||||
|
||||
LLM 在调用 recall_memory 前会将精确查询泛化:
|
||||
- "postgres v15.3 + pgvector 0.5.1" → "PostgreSQL 数据库"
|
||||
- 版本号、扩展名、错误码全部丢失
|
||||
- BM25 匹配不到精确 token → 词法腿报废
|
||||
|
||||
## 方案
|
||||
|
||||
在 `messages.append(user_msg)` 和 LLM 调用之间插入:
|
||||
1. 用 `user_message` 原串调用 recall_memory
|
||||
2. 结果以合成 assistant tool_call 注入对话历史
|
||||
3. LLM 后续调用时同时看到用户消息和 pre-recall 结果
|
||||
|
||||
## 为什么重要
|
||||
|
||||
Ablation 证实 query expansion(LLM paraphrase)反而降低性能——BM25 已捕获精确 token,dense 已捕获语义改写,额外 LLM 改写层引入噪音。
|
||||
|
||||
## 参考
|
||||
- [[atlas-agent-memory-architecture-2026]]
|
||||
- [[hybrid-recall-pipeline]]
|
||||
- [[bm25-financial-retrieval]]
|
||||
Reference in New Issue
Block a user