How We Built a Searchable Knowledge Base for a Physical Security Operations Center

A physical security operations center manages a constant flow of procedures, protocols, incident types, and regulatory requirements. Most of that knowledge lives in PDFs and Word documents that operators can’t search quickly during an active incident. We built a hybrid full-text and semantic search system that makes an entire SOP library queryable in milliseconds.

The Starting Point

The operations center had 143 operational documents: standard operating procedures, quick reference guides, incident response protocols, equipment manuals, and regulatory compliance guides. The documents ranged from 2 to 80 pages. Operators knew what existed but couldn’t quickly locate the specific paragraph they needed during a fast-moving situation. The answer to most questions was “check the SOP folder”—which wasn’t fast enough.

The Indexing Pipeline

Each document was extracted to plain text, chunked into 400-token segments with 50-token overlap (so context isn’t lost at chunk boundaries), and indexed two ways: into SQLite FTS5 for keyword search, and into a vector table using locally-generated embeddings for semantic search.

The result: 4,232 searchable chunks across all 143 documents, stored in a 27MB SQLite file. No cloud services. No API keys. No running server required.

The Search Interface

Operators can search in natural language: “what do I do when a visitor tailgates through a secured door?” or “fire evacuation procedure floor 3.” The query runs against both indexes simultaneously. FTS5 handles exact keyword matches and Boolean operators; the vector index surfaces semantically related content even when the exact words don’t appear. Results are merged and ranked with the most relevant chunk shown first.

Each result shows the source document, the relevant excerpt, and a relevance score. Operators can pull the full document section with one additional query.

AI-Powered Micro-SOPs

An additional layer generates micro-SOPs on demand: condensed, step-by-step summaries of procedures extracted from the full documents. When an operator queries a complex procedure, the system can return both the raw source text and a simplified, checklist-format version generated from it. This bridges the gap between comprehensive documentation and fast execution.

Deployment

The system was deployed as a static web application that any operator can open from any device on the network. The search backend is a lightweight Python server. Total deployment time from indexed documents to working search interface was under two hours. The entire system, including all 143 documents’ worth of indexed content, fits in a single directory.

Impact

The most significant change was reducing the time to locate a specific procedure from “find the right PDF, open it, Ctrl+F for a keyword, navigate to the right section” to a single natural-language query returning the relevant paragraph in under a second. For a high-pressure environment where seconds matter, that’s a meaningful operational improvement.