The RAG-Ready Vector Database

A vector database indexes embeddings of shots, diagnostics, and documents so copilots retrieve the most relevant context by meaning, not keyword.

Retrieval by meaning

The copilots reason over a large corpus: past shots, diagnostic traces, engineering documents, procedures, and the schema registry. A vector database stores embeddings of these items so a query is answered by semantic similarity — the copilot retrieves the shots and documents closest in meaning, then reasons over them. This is the retrieval half of retrieval-augmented generation.

What is embedded

• Diagnostic and shot summaries, so 'find shots with a growing n=1 mode near a density limit' returns the right ones.
• Engineering and procedure documents, for the operations and maintenance copilots.
• Schema and registry entries, so a copilot can ground itself in the current channel and feature definitions.
• Machine-scoped: breeder and burner corpora are indexed and queryable separately or together.

How retrieval works

import numpy as np
def top_k(query_vec, index, k=5):
    # cosine similarity retrieval over normalized embeddings
    sims = index @ (query_vec/np.linalg.norm(query_vec))
    return np.argsort(-sims)[:k]

Grounded and governed

Retrieval keeps the copilots grounded in the plant's real data rather than free-associating. Every retrieved item carries its lineage, so a copilot's answer can be traced to the exact shots and documents it rested on — essential for a safety-relevant system. The vector database is built from the same governed corpus as the archive and registry (see document & schema registry), and it is a design specification for the AI stack serving both machines.