docs: update sequence diagram, memory flow, vectorization strategy for StaticEmbedder + CleanTemplateText + three-branch vector

2026-09-24 02:48:04 +00:00 · 2026-07-17 21:51:03 +08:00
parent 3c441649c9
commit f9ea80abbf
6 changed files with 104 additions and 94 deletions
--- a/docs/en/ARCHITECTURE.md
+++ b/docs/en/ARCHITECTURE.md
@ -75,42 +75,44 @@ Setting `ctx.Response` at any stage jumps to `after_output`.
 ```
 ① Context (Working Window)
   RelevanceContext — In-memory events[] + JSON persistence
-   Append: Each input, Vectorize(char 1-2gram TF-IDF)
-   Prune:  TF-IDF CosineSimilarity, keep topK + last 10
-       ├── Keep → timeline → system prompt (sorted by time)
-       └── Low score → Document layer archive (ContextToDoc)
+   Append: Each input, CleanTemplateText → three-branch vector(textForVector)
+           agent→Response, user→Input, cold_storage→Input+Response
+           StaticEmbedder pretrained word embedding / TF-IDF fallback
+   Prune:  StaticEmbedder CosineSimilarity, keep topK + last 10
+       ├── Keep → timeline → chronologically sorted → system prompt
+       └── Low score → Document layer archive (original timestamp)
   Save: 5s debounce write to disk

-        ↓ Prune archive                       ↑ LLM active recall
+         ↓ Prune archive                       ↑ LLM active recall

 ② Document (File Memory)
   DocStore — JSON files + TF-IDF InvertedIndex
   Write: Prune archive / doc_commit / Graph snapshot (syncGraphToDocs)
   Read:
-       ├── Auto-inject: Query(input, top3) → [Related Memory Docs] → system prompt (read-only, update AccessCount)
+       ├── Auto-inject: Query(input, top3) → similarity summary → [Related Memory Docs] → system prompt (read-only)
       └── LLM active:  doc_query → Consume(read and delete)
                         → context.Append{Timestamp: d.CreatedAt, Source: "cold_storage"} per doc
                         → Docs written to context timeline with original timestamps, deleted from docStore
   Cold: FindColdDocs(72h, ≤2 accesses) → docToTriples → Graph

-        ↓ Cold doc distillation               ↑ Auto recall
+         ↓ Cold doc distillation               ↑ Auto recall

 ③ Graph (Graph Database)
   SQLite — entities + relations tables
-   Write: memory_commit / cold doc distillation / Pipeline rule distillation
+   Write: memory_commit / cold doc distillation / Pipeline rule distillation / memory_merge
   Read:
       ├── Auto recall: Indexer.BuildContext(input)
-       │     → TF-IDF entity name search → BFS depth=2
+       │     → CleanTemplateText → vector entity search + jieba keywords → SQLite LIKE + BFS depth=2
       │     → [Memory Index] → system prompt
-       └── LLM active: memory_recall / doc_query
+       └── LLM active: memory_recall / memory_merge / memory_purge / memory_edit / memory_delete_entity
   Social: person_query / set_trait / relate (wraps GraphDB)

 ④ Distillation Pipeline (30min heartbeat)
   distillContext → window > 2×maxSize → force Prune
   syncGraphToDocs → Graph snapshot to Document (cross-layer searchable)
   reorgGraph:
-       Step1: indexer.Sync — rebuild entity TF-IDF vector index
-       Step2: docStore.Reindex — rebuild document TF-IDF vector index
+       Step1: indexer.Sync — rebuild entity vector index
+       Step2: docStore.Reindex — rebuild document vector index
       Step3: Cold docs → docToTriples → GraphDB.Commit
       Step4: Entity similarity (Bigram Jaccard > 0.75) → consolidation → LLM decides merge
       Step5: evaluateGraphQuality → LLM decides keep/delete
@ -121,36 +123,35 @@ Setting `ctx.Response` at any stage jumps to `after_output`.
     → triples → GraphDB.Commit
 ```

-### Vectorization: Two Strategies
+### Vectorization: Pretrained Word Embedding + TF-IDF Fallback

-Vectorization is used in 4 independent locations with different strategies:
+All vectorization unified under `StaticEmbedder` (`internal/memory/static_embedder.go`):

-**Strategy A — Local Word Embedding** (`LocalWordEmbedder`, `internal/memory/embedder.go`), used by Context layer:
-
- **jieba tokenization** → removes stop words and single characters
- **TF-IDF** as base word weight
- **Sliding window (size=5)** counts word co-occurrence → **PMI (Pointwise Mutual Information)** → keeps top 50
- **Vectorization**: `vec[ctx] += TF-IDF × PMI` + self-tag `__w__` + TF-IDF
-
-**Strategy B — char-bigram TF-IDF + jieba keyword extraction** (`TFIDFVectorizer` + `ExtractKeywords`), used by Document and Indexer layers:
-
- **char bigram tokenization** (1-2 gram) for entity name vector search
- **jieba tokenization** for keyword extraction, paired with SQLite LIKE + BFS traversal
- **TF-IDF weights** + **inverted index**
+**Primary Strategy — Pretrained Word Embedding (aligned 300d)**
+- Model sources: ConceptNet Numberbatch (77-language aligned) / fastText Chinese / fastText English
+- Configured via `core.agent.embedding_model_path` (comma-separated multi-model)
+- Path containing `numberbatch` → auto-download ConceptNet; `cc.zh.` → fastText Chinese; `cc.en.` → fastText English
+- Falls back to ConceptNet by default if no match
+- **Pre-processing**: `CleanTemplateText` strips QQ tool-call templates and timestamp noise
+- **Three-branch vector source**: agent→Response, user→Input, cold_storage→Input+Response
+- **TF-IDF fallback**: auto-fallback to bag-of-words TF-IDF if model download fails or not configured

 | Location | File | Purpose | Algorithm |
 |----------|------|---------|-----------|
-| Context Prune | `context.go:161` | Trim low-relevance context events | LocalWordEmbedder → CosineSimilarity(queryVec, evt.Vector) |
-| DocStore Query | `document.go:198` | Recall related content from document memory | char-bigram TF-IDF + jieba keywords → InvertedIndex + CosineSimilarity |
-| Indexer Entity Search | `indexer.go:149` | Recall related entities from Graph | char-bigram TF-IDF vector search + jieba keywords → InvertedIndex + CosineSimilarity + SQLite BFS |
-| Entity Similarity Detection | `agent.go:2297` | Detect similar entities in Graph | Bigram Jaccard (>0.75 → consolidation) |
+| Context Prune | `context.go:155` | Trim low-relevance context events | VectorizeClean → CosineSimilarity(queryVec, evt.Vector) |
+| DocStore Query | `document.go:206` | Recall from document memory | TF-IDF Vectorize → vec.Search |
+| Indexer Entity Search | `indexer.go:96+111` | Recall from Graph | vector entity search + jieba keywords → SQLite LIKE + BFS |
+| Entity Similarity Detection | `agent.go` | Detect similar entities in Graph | Bigram Jaccard (>0.75 → consolidation) |

 ### Context Layer

 `internal/agent/core/context.go` — `RelevanceContext`
 - Maintains recent event list, writes JSON on each Append/Prune to prevent data loss
- Word embedding relevance scoring on user input (LocalWordEmbedder → CosineSimilarity), keeps topK
+- Pre-vectorization pipeline runs through `CleanTemplateText` to remove template noise
+- Three-branch `textForVector`: agent events → Response, user events → Input, cold_storage → Input+Response
+- Pretrained word embedding `StaticEmbedder` → CosineSimilarity, auto-fallback to TF-IDF if unavailable
 - Protects last 10 events from eviction; excess candidates are sorted by relevance and archived to document memory
+- Archived events retain original timestamps; on `doc_query` recall they re-insert into the context timeline at their original position

 ### Document Layer

@ -171,6 +172,10 @@ Vectorization is used in 4 independent locations with different strategies:
 |------|---------|
 | `memory_recall` | Recall from Graph |
 | `memory_commit` | Write triples to Graph |
+| `memory_merge` | Merge two entity nodes |
+| `memory_purge` | Delete entity node |
+| `memory_edit` | Edit existing entity/relation |
+| `memory_delete_entity` | Delete entity and all its relations |
 | `memory_introspect` | View memory statistics |
 | `doc_query` | Search from Document |
 | `doc_commit` | Write to Document |
--- a/docs/en/OVERVIEW.md
+++ b/docs/en/OVERVIEW.md
@ -37,7 +37,7 @@ Code is in the project root, implemented in Go.
 - Maintains a message loop (`eventLoop`), queuing input from the IO layer
 - Each input goes through the full processing pipeline: memory recall → persona injection → LLM call → tool execution → output delivery
 - LLM calls abstracted through Provider interface, supports 8 LLM sources with automatic fallback
- Context management (`context.go`) based on word embedding scoring (LocalWordEmbedder → CosineSimilarity), automatic pruning of low-relevance events
+- Context management (`context.go`) based on pretrained word embedding scoring (StaticEmbedder → CosineSimilarity, TF-IDF fallback), automatic pruning of low-relevance events

 **Memory System** (`internal/memory/`):
 - **GraphDB** (`graph.go`) — SQLite, entities + relations tables, BFS traversal
--- a/docs/zh/ARCHITECTURE.md
+++ b/docs/zh/ARCHITECTURE.md
@ -75,42 +75,44 @@ eventLoop() → processTextInput()
 ```
 ① Context (工作窗口)
   RelevanceContext — 内存 events[] + JSON持久化
-   Append: 每次输入, Vectorize(char 1-2gram TF-IDF)
-   Prune:  TF-IDF CosineSimilarity, 保留 topK + 最近10条
-       ├── 保留 → timeline → system prompt (按时间排序)
-       └── 低分 → Document 层归档 (ContextToDoc)
+   Append: 每次输入, CleanTemplateText → 三分支向量(textForVector)
+            agent事件→Response, 用户事件→Input, cold_storage→Input+Response
+            StaticEmbedder 预训练词嵌入 / TF-IDF 回退
+   Prune:  StaticEmbedder CosineSimilarity, 保留 topK + 最近10条
+       ├── 保留 → timeline → 按时间排序 → system prompt
+       └── 低分 → Document 层归档 (原始时间戳)
   Save: 5s debounce 写盘

-        ↓ Prune 归档                          ↑ LLM 主动召回
+         ↓ Prune 归档                          ↑ LLM 主动召回

 ② Document (文件记忆)
   DocStore — JSON文件 + TF-IDF InvertedIndex
   写入: Prune归档 / doc_commit / Graph快照(syncGraphToDocs)
   读取:
-       ├── 自动注入: Query(input, top3) → 【相关记忆文档】→ system prompt (只读, 更新 AccessCount)
+       ├── 自动注入: Query(input, top3) → 相似度摘要 → 【相关记忆文档】→ system prompt (只读)
       └── LLM主动:  doc_query → Consume(读取并删除)
                         → 逐条 context.Append{Timestamp: d.CreatedAt, Source: "cold_storage"}
                         → 文档以原始时间戳写入 context 时间线, 从 docStore 删除
   冷化: FindColdDocs(72h, ≤2次访问) → docToTriples → Graph

-        ↓ 冷文档蒸馏                           ↑ 自动召回
+         ↓ 冷文档蒸馏                           ↑ 自动召回

 ③ Graph (图数据库)
   SQLite — entities + relations 表
-   写入: memory_commit / 冷文档蒸馏 / Pipeline 规则蒸馏
+   写入: memory_commit / 冷文档蒸馏 / Pipeline 规则蒸馏 / memory_merge
   读取:
       ├── 自动召回: Indexer.BuildContext(input)
-       │     → TF-IDF 实体名搜索 → BFS depth=2
+       │     → CleanTemplateText → 向量实体搜索 + jieba关键词 → SQLite LIKE + BFS depth=2
       │     → 【记忆索引】→ system prompt
-       └── LLM主动: memory_recall / doc_query
+       └── LLM主动: memory_recall / memory_merge / memory_purge / memory_edit / memory_delete_entity
   Social: person_query / set_trait / relate (包装 GraphDB)

 ④ 蒸馏管道 (每30min心跳)
   distillContext → 窗口>2×maxSize → 强制Prune
   syncGraphToDocs → Graph 快照写入 Document(跨层可搜索)
   reorgGraph:
-       Step1: indexer.Sync — 重建实体TF-IDF向量索引
-       Step2: docStore.Reindex — 重建文档TF-IDF向量索引
+       Step1: indexer.Sync — 重建实体向量索引
+       Step2: docStore.Reindex — 重建文档向量索引
       Step3: 冷文档 → docToTriples → GraphDB.Commit
       Step4: 实体相似度(Bigram Jaccard>0.75) → consolidation → LLM判断合并
       Step5: evaluateGraphQuality → LLM判断保留/删除
@ -121,36 +123,35 @@ eventLoop() → processTextInput()
     → 三元组 → GraphDB.Commit
 ```

-### 向量化算法：两种策略
+### 向量化：预训练词嵌入 + TF-IDF 回退

-向量化在 4 个独立位置以不同方式使用：
+所有向量化统一使用 `StaticEmbedder`（`internal/memory/static_embedder.go`）：

-**策略 A — 局部词嵌入**（`LocalWordEmbedder`, `internal/memory/embedder.go`），用于 Context 层：
-
- **jieba 分词** → 去除停用词和单字
- **TF-IDF** 作为基础词权重
- **滑动窗口（size=5）** 统计词对共现 → **PMI（点互信息）** → 保留 top 50
- **向量化**：`vec[ctx] += TF-IDF × PMI` + 自身上标 `__w__` + TF-IDF
-
-**策略 B — char-bigram TF-IDF + jieba 关键词提取**（`TFIDFVectorizer` + `ExtractKeywords`），用于 Document 和 Indexer 层：
-
- **char bigram 分词**（1-2 gram）用于实体名向量搜索
- **jieba 分词**用于关键词提取，配合 SQLite LIKE + BFS 遍历
- **TF-IDF 权重** + **倒排索引**
+**主策略 — 预训练词嵌入（词对齐 300 维）**
+- 模型来源：ConceptNet Numberbatch（77 语对齐）/ fastText 中文 / fastText 英文
+- 通过 `core.agent.embedding_model_path` 配置（逗号分隔多模型）
+- 路径名含 `numberbatch` → 自动下载 ConceptNet，含 `cc.zh.` → fastText 中文，含 `cc.en.` → fastText 英文
+- 不匹配则默认 ConceptNet
+- **前处理**：`CleanTemplateText` 剥离 QQ 工具调用模版、时间戳噪声，避免垃圾干扰相似度
+- **三分支向量来源**：agent→Response，用户→Input，cold_storage→Input+Response
+- **TF-IDF 回退**：模型下载失败或未配置时自动回退词袋 TF-IDF，服务不中断

 | 位置 | 文件 | 用途 | 算法 |
 |------|------|------|------|
-| Context Prune | `context.go:161` | 裁剪低相关性上下文事件 | LocalWordEmbedder → CosineSimilarity(queryVec, evt.Vector) |
-| DocStore Query | `document.go:198` | 从文档记忆召回相关内容 | char-bigram TF-IDF + jieba 关键词 → InvertedIndex + CosineSimilarity |
-| Indexer 实体搜索 | `indexer.go:149` | 从Graph召回相关实体 | char-bigram TF-IDF 向量搜索 + jieba 关键词 → InvertedIndex + CosineSimilarity + SQLite BFS |
-| 实体相似度检测 | `agent.go:2297` | 检测Graph中相似实体 | Bigram Jaccard (>0.75 → consolidation) |
+| Context Prune | `context.go:155` | 裁剪低相关性上下文事件 | VectorizeClean → CosineSimilarity(queryVec, evt.Vector) |
+| DocStore Query | `document.go:206` | 文档记忆召回 | TF-IDF Vectorize → vec.Search |
+| Indexer 实体搜索 | `indexer.go:96+111` | Graph实体召回 | 向量实体搜索 + jieba关键词 → SQLite LIKE + BFS |
+| 实体相似度检测 | `agent.go` | Graph中相似实体 | Bigram Jaccard (>0.75 → consolidation) |

 ### Context 层

 `internal/agent/core/context.go` — `RelevanceContext`
 - 维护最近事件列表，每次 Append/Prune 写入 JSON 防丢
- 用户输入时做词嵌入相关性评分（LocalWordEmbedder → CosineSimilarity），保留 topK
+- 向量化前统一经 `CleanTemplateText` 去模版噪声
+- 三分支 `textForVector`：agent 事件用 Response、用户事件用 Input、cold_storage 用 Input+Response
+- 预训练词嵌入 `StaticEmbedder` → CosineSimilarity，模型不可用时自动回退 TF-IDF
 - 保护最近 10 条记录免于淘汰，超出部分按相关性排序归档到文档记忆
+- 归档事件以原始时间戳写入文档记忆，后续 `doc_query` 召回时按原始时间戳插回时序

 ### Document 层

@ -171,6 +172,10 @@ eventLoop() → processTextInput()
 |------|------|
 | `memory_recall` | 从 Graph 召回 |
 | `memory_commit` | 写入 Graph 三元组 |
+| `memory_merge` | 合并两个实体节点 |
+| `memory_purge` | 删除指定实体 |
+| `memory_edit` | 编辑已有实体/关系 |
+| `memory_delete_entity` | 删除实体及其所有关系 |
 | `memory_introspect` | 查看记忆统计 |
 | `doc_query` | 从 Document 搜索 |
 | `doc_commit` | 写入 Document |
--- a/docs/zh/OVERVIEW.md
+++ b/docs/zh/OVERVIEW.md
@ -37,7 +37,7 @@ HomeAgent 是一个持续运行的个人智能 Agent 框架。
 - 维护一个消息循环（`eventLoop`），从 IO 层排队接收输入
 - 每次输入走完整的处理管道：记忆召回 → 人格注入 → LLM 调用 → 工具执行 → 输出发送
 - LLM 调用通过 Provider 接口抽象，支持 8 个 LLM 源自动降级
- 上下文管理（`context.go`）基于词嵌入评分（LocalWordEmbedder → CosineSimilarity），自动剪枝低相关性事件
+- 上下文管理（`context.go`）基于预训练词嵌入评分（StaticEmbedder → CosineSimilarity，TF-IDF回退），自动剪枝低相关性事件

 **记忆系统** (`internal/memory/`)：
 - **GraphDB** (`graph.go`) — SQLite，entities + relations 表，BFS 遍历