WECHASER/ 技术架构TECH ARCHITECTURE
// TECH DEEP-DIVE

技术架构。 Tech architecture.

没有市场词。这页讲清楚 WeChaser / 时光脉络的代码、数据、调用链是怎么搭起来的——技术决策人可以直接判断这套东西能不能进你机房。 No marketing words. This page lays out how WeChaser / Shiguangmailuo is actually built — code, data, call paths — so technical decision-makers can judge directly whether it fits your stack.

01 / 系统全貌System overview

两大核心系统：
Hermes Agent + 时光脉络。 Two core systems:
Hermes Agent + Shiguangmailuo.

Hermes 是神经系统（推理 / 调度 / 记忆），时光脉络是知识躯体（图谱 / 向量 / 文档）。中间走 MCP Bridge。 Hermes is the nervous system (reasoning / scheduling / memory). Shiguangmailuo is the knowledge body (graph / vector / docs). MCP Bridge connects them.
用户交互层 / USER LAYERUSER LAYER
Lark BotFEISHU
Web UIGRADIO
REST APIFASTAPI
CLI / MCPOPENCLAW
// MCP / HTTP / STDIO
核心系统 / CORECORE SYSTEMS
Hermes Agent
神经系统NERVOUS
run_agent.py · memory · cron · skills · gateway
时光脉络 APIShiguangmailuo API
知识躯体BODY
FastAPI · LangGraph · ingest_pipeline · hybrid_retriever
数据存储层 / STORAGESTORAGE LAYER
Neo4jGRAPH
QdrantVECTORS
MySQLHISTORY
SQLiteSKILLS · STATE
02 / 数据流Data flow

入库 + 查询 + 进化，
三条独立 pipeline。 Ingest + Query + Evolve,
three independent pipelines.

入库 / IngestIngest

01

格式检测Format detection

FileParser · PDF / Word / Excel / PPT / Image

02

解析分块Parse + chunk

DocumentParser · MinerU · PaddleOCR

03

LLM 结构化抽取LLM extract

DeepSeek-Reasoner / GLM-4.5 · prompts/ 全部专门设计

04

数据验证Validate

Validator · file_hash 唯一约束 · schema check

05

图谱入库 + 向量入库Graph + vector load

Neo4jLoader · Qdrant VectorStore · 1024-d embedding

查询 / QueryQuery

01

意图识别 + 路由Intent route

quick / standard / deep · 不同档位走不同 pipeline

02

三路混合检索Hybrid retrieve

vector + graph + sparse · 并行

03

RRF 融合 + MMR 多样性RRF + MMR

rerank fusion + redundancy removal

04

BGE 精排BGE rerank

BAAI/bge-reranker-base · 最终 top-k

05

LLM 综合 + 溯源Synthesis + cite

DeepSeek-Chat · grounded RAG · 100% citation

03 / 技术栈Stack

每一项都有"能跑在客户机房"
的兜底版本。 Every choice has
"runs on customer hardware" as a constraint.

LAYER
主选 / 备选PRIMARY / FALLBACK
说明NOTE
API
FastAPI · port 1000 · Uvicorn
同步 + WebSocket。Sync + WS.
Web UI
Gradio · port 7860 + Ingest UI :7861
客户内部使用。Internal admin.
Agent
LangGraph · agent_graph/orchestrator.py
90 轮工具调用上限。90-turn cap.
Graph DB
Neo4j 5.15+ · 534 nodes · file_hash unique
核心查询主体。Primary store.
Vector
Qdrant · 890 points · cosine
Milvus 备选。Milvus fallback.
Embedding
DashScope text-embedding-v4 · 1024-d
本地化时切 BGE。BGE for offline.
Reranker
BGE Reranker · BAAI/bge-reranker-base
本地推理。Local infer.
LLM (extract)
DeepSeek-Reasoner · DeepSeek-Chat · Qwen
结构化抽取主用。Structured.
LLM (multimodal)
Gemini 2.5 Flash/Pro · GLM-4.5 · GPT-4o
原理图 / Word / 图像。Schematic / image.
OCR
PaddleOCR + LibreOffice fallback
本地，离线。Local, offline.
Optimization
OR-Tools CP-SAT
BOM 选型求解。BOM selection.
History
MySQL · 会话 + feedback
SQLite 单机替代。SQLite swap.
Deployment
docker-compose · on-prem · air-gap
客户机房一键起。One-click on customer host.
04 / MCP 工具集MCP tools

已交付 4 个工具。
规划中 8 个。 4 tools shipped.
8 more in plan.

检索Retrieve
kg_search(query, top_k)
三路混合检索 → 融合 → 精排，返回结构化原料 + 引用。Hybrid 3-way retrieval → fusion → rerank, returns structured material + citations.
SHIPPED
检索Retrieve
kg_graph_query(cypher)
原生 Cypher 查询。Agent 直接对图谱发 query 用。Raw Cypher query for direct graph access by the Agent.
SHIPPED
检索Retrieve
kg_board_query(board_id)
板卡完整信息：参数 + 元器件 + 文档 + 关系图。Full board info: parameters + components + docs + relations.
SHIPPED
反馈Feedback
kg_add_feedback(qid, score)
用户反馈记录到 reflection 日志，用于自修补和精炼。Records user feedback into the reflection log for self-patching and refinement.
SHIPPED
检索Retrieve
kg_find_component(...)
查找元器件 + 替代关系 + 兼容性。Find component + replacement + compatibility relations.
TODO
写入Write
kg_patch_gap(...)
图谱自修补：检测语义有图谱无的情况，生成补丁 → 审核队列。Auto-patch: detects "vector has it, graph doesn't" gap, generates patch → review queue.
TODO
入库Ingest
kg_ingest_document(...)
触发文档摄入 pipeline。Agent 直接喂文档。Triggers ingest pipeline. Agent feeds docs directly.
TODO
精炼Synthesis
kg_synthesize(topic)
跨文档精炼：发现"最近 3 份报告都提到 X"这种洞察。Cross-doc synthesis: surfaces patterns like "last 3 reports all mention X".
TODO
05 / 设计原则Design principles

五条铁律，
所有技术决策的尺子。 Five iron laws —
the ruler we measure decisions against.

AI 准确度 = 用户表达 × AI 理解度Accuracy = user expression × AI understanding

提升整体准确度有两条路：让用户表达更精确（不可控），或让 AI 更懂这家企业（我们的事）。后者是知识库的核心价值。

To raise accuracy you can either tighten user expression (out of our hands) or deepen AI's understanding of this specific company (our job). The latter is the entire value of a knowledge base.

impl hybrid retrieval + grounded RAG + reflection loop

知识库不直接解决问题The KB doesn't solve problems directly

它让其他 AI 工具变得能解决问题。边界要清晰——我们不做生成，不抢调用方的活。这条原则保护我们不被功能蔓延拖垮。

It makes other AI tools capable of solving them. The boundary stays sharp — we don't generate, we don't compete with the calling agent. This is what protects us from feature creep.

impl middleware-only · MCP / CLI / REST as the only surface

生成 ≠ 检索Generation ≠ retrieval

知识库只做后者，前者交给调用方 Agent。这避免了"知识库幻觉"——大多数 RAG 灾难都来自把生成揉进检索环节。

The KB does only retrieval; generation belongs to the calling agent. This avoids the "knowledge-base hallucination" failure mode — most RAG disasters come from blending generation into retrieval.

impl structured facts + citations + confidence — never prose

工业级零容忍幻觉Industrial-grade zero-hallucination

未来要介入工业生产决策。决策辅助系统级别的责任 ≠ 聊天玩具级别的容忍度。100% 溯源是底线，不知道就拒答。

We are headed toward supporting industrial production decisions. Decision-support liability ≠ chat-toy tolerance. 100% citation is the floor; if we don't know, we refuse.

impl citation-required output schema · refusal pathway built in

先用起来，再迭代Run first, iterate after

不被花哨 UI 或未来场景拖住。第一阶段：能简单安装、操作、有 app 图标、有 WebUI。等真用起来，反馈来推迭代方向。

Don't be paralysed by polished UI or hypothetical scenarios. Phase one: install, operate, app icon, WebUI. Let real usage drive what to build next.

impl shipped early · feedback-driven roadmap · weekly review
更深入？Deeper?

需要更具体的
代码 / 部署方案 ？ Want a deeper look at
code / deployment?

如果你是技术负责人，我们可以走更深的会议——演示 MCP 接入流程、跑一遍 docker-compose 到客户机房的部署。 If you're a technical lead, we can do a deeper session — live demo of MCP integration and a docker-compose deployment dry-run on your hardware.

预约技术演示Book tech demo→ 产品页Product page→