企业 AI 生态的中间件——让其他 AI 工具变"懂企业"。不直接面对终端用户，不和 ChatGPT 竞争。是有灵魂的图书馆 + 能主动学习的管理 Agent。 Middleware for the enterprise AI ecosystem — making other AI tools actually understand a specific company. We don't face end users; we don't compete with ChatGPT. We're a library with a soul, plus a self-learning manager agent.
用户不直接和知识库对话。其他 Agent—— OpenClaw / 飞书 AI / 钉钉 AI / Claude / 自研 Bot ——通过 MCP / CLI / REST 调用，拿到检索片段、结构化原料、文件超链接。 Users don't talk to the knowledge base directly. Other agents — OpenClaw / Lark AI / DingTalk AI / Claude / your in-house bot — call us via MCP / CLI / REST and get retrieval snippets, structured material, and file links.
生成（写营销话术 / 写产品规划书 / 设计海报）留给调用方 Agent。文件原件留在你本地服务器，永不外传。 Generation — marketing copy, planning docs, posters — stays with the calling agent. File originals stay on your local server, never leaving the perimeter.
这是个反 SaaS 选择。不卖通用聊天框，不抢用户注意力。我们的客户是企业，不是消费者。 This is an anti-SaaS choice. We don't sell a generic chatbox or compete for user attention. Our customer is the enterprise, not the consumer.
把复合问题拆成可查询的子问题。
"X 产品支持哪些接口、用了什么芯片、有没有兼容问题？"——这一句话至少包含三个独立检索意图：协议匹配 / 元器件清单 / 冲突关系。每个子问题走最适合它的检索路径，最后聚合。
Compound questions become atomic sub-queries.
"What interfaces does X support, which chip, any compat issues?" hides at least three retrieval intents: protocol match / component listing / conflict graph. Each routes to its own path, then aggregates.
三路混合检索：向量库 + 知识图谱 + 全文索引。
RRF 融合排序，MMR 多样性过滤，BGE Reranker 精排。意图识别后路由——快速 / 标准 / 深度三档，不浪费算力，不漏召回。
Hybrid retrieval: vector store + knowledge graph + full-text.
RRF fusion, MMR diversity, BGE Reranker for final ordering. Intent-routed: quick / standard / deep — no wasted compute, no missed recall.
检索片段去重、矛盾标识、置信度打分，组装成结构化原料。
调用方 Agent 拿到的不是一团文字，而是 {facts: [...], citations: [...], confidence: 0.87, gaps: [...]}。每条原料都有溯源链接，不知道就明确说"信息不足"。
Dedup, conflict-flag, confidence-score — assemble into structured material.
Calling agent receives {facts: [...], citations: [...], confidence: 0.87, gaps: [...]} — not a wall of text. Every fact carries a source. When unsure, we say so explicitly.
每句话都有出处。无出处的话不许说。
Every sentence has a source. No source = no statement.
不知道就明确说"信息不足，需要补充 X 类文档"。
If we don't know, we say "insufficient information — add docs of type X".
找近似时主动告诉调用方："这是相关但不完全匹配的内容，置信度 0.62"。
For near-misses we surface the score: "related, not exact, confidence 0.62".
大模型训练数据中类似产品的参数 ≠ 此产品的参数。一律不许生造。
Numbers from the model's training corpus for "similar" products are not this product's numbers. No invention.
把 A 产品规格说成 B 产品。在电子硬件场景下这是事故级错误。
Reporting product A's spec as product B's. In hardware that's an incident-level error.
引用对的文件但理解错了原文。这种错误最危险——表面有出处。
Cite the right doc, get the meaning wrong. The most dangerous form — the citation looks valid.
优先方式。Claude Code、Hermes Agent、其他 MCP-aware 工具直接挂接。已通过 313/319 stdio 集成测试。
Primary path. Claude Code, Hermes Agent, other MCP-aware tools plug in directly. 313/319 stdio integration tests passing.
// 4 tools exposed kg_search(query, top_k) kg_graph_query(cypher) kg_board_query(board_id) kg_add_feedback(qid, score)
调用方 Agent 不支持 MCP？走 CLI——子进程标准输入输出，跨语言通用。OpenCLI 适配器已就绪。
Calling agent doesn't speak MCP? Use CLI — subprocess stdin/stdout, language-agnostic. OpenCLI adapter ready.
$ shiguang search "主控板接口列表" → facts: [3], citations: [3], confidence: 0.91
FastAPI 暴露的标准 HTTP 端点，端口 1000。前端 / 浏览器扩展 / 第三方系统直连。
FastAPI HTTP endpoints on port 1000. Frontends, browser extensions, third-party systems plug in.
POST /api/v1/retrieve { "query": "...", "mode": "hybrid", "top_k": 10 }
30 分钟通话，我们带着可行性评估表上门。第一阶段调研免费。 A 30-minute call. We bring a feasibility checklist. Phase one is free.