Knowledge Base — Manufacturing Search Engine

A precision information-retrieval service for manufacturing knowledge. Built on Elasticsearch with hybrid BM25 + vector search — not RAG, not generative AI. Documents are returned verbatim or not at all.

Why not RAG?

In manufacturing, alarm codes differ by one character, equipment parameters are meaningless without domain context, and wrong answers have real consequences. This system is designed around a zero-fabrication guarantee: if a document matches, it is shown as-is; if nothing matches, the caller is told so explicitly. The LLM is used only as a query-understanding proxy and a conversational explainer — never as a source of facts.

---

What this project is

A zero-fabrication knowledge-base API for semiconductor manufacturing equipment. Documents are retrieved verbatim from Elasticsearch; the system never generates document text. The LLM appears in exactly two roles:

• Query-understanding proxy (POST /api/v1/extract) — turns a free-text question into structured search parameters.
• Conversational search assistant (POST /api/v1/chat) — extracts params, searches the KB, and answers strictly from the returned documents.

Stack: FastAPI · Elasticsearch 8.x (IK analyzer plugin) · pydantic-settings · httpx · DashScope Embeddings API (optional).

---

The three guarantees

• Zero fabrication

  Search responses contain verbatim document sections or nothing. The LLM is forbidden from inventing parameters, steps, or alarm codes.

• Graceful degradation

  No LLM key → search and indexing still work (AI chat returns 503). No embedding key → keyword-only BM25 search, no kNN. The server always boots.

• Taxonomy enforcement

  project and equipment values are validated against config/taxonomy.yaml at index time. Unknown values are rejected rather than silently stored.

---

Where to go next

If you want to…	Read
Run the stack and try it out	Getting Started
Understand the big picture	Architecture → Overview
Understand the chat/extract endpoints	Architecture → AI Chat Search
Understand how files become documents	Architecture → Import Pipeline
Understand the search ranking & status contract	Architecture → Search & Ranking
Tune settings / add a provider	Configuration
Call the HTTP API	API Reference
Wire up metrics & logs	Observability
Rebuild the system from scratch	Reference → Build from Scratch
Look up every field & ES mapping	Reference → Data Model
Look up every setting & env var	Reference → Configuration Reference
Deploy, back up & harden	Operations → Deployment
Run the tests / develop	Operations → Testing & Development
Diagnose a failure	Operations → Troubleshooting
Understand the security posture	Operations → Security

---

At a glance

flowchart TD
    UI["Upstream chat / UI layer"] -->|SearchRequest| API
    subgraph API["FastAPI (kb.main)"]
        search["POST /api/v1/search"]
        chat["POST /api/v1/chat · /extract"]
        ingest["POST /api/v1/ingest/*"]
        docs["POST /api/v1/documents/*"]
    end
    API --> ES[("Elasticsearch 8.x<br/>BM25 + kNN")]
    API -.optional.-> EMB["DashScope Embeddings<br/>text-embedding-v3 (1024-dim)"]
    API -.optional.-> LLM["LLM (OpenAI-compatible)<br/>extract · segment · explain"]

Structured filters narrow the candidate set first, then hybrid BM25 keyword search and dense-vector similarity re-rank results. The caller never sees AI-generated text — only verbatim document sections.