AI code search over a whole business system
A UK procurement business with a .NET API, a WordPress site and a customised Zoho CRM
A search engine that gives an AI coding assistant cited answers about a business's code, CRM rules and three years of history, in about three seconds.
The problem
The business runs on a .NET Web API, a WordPress site, Zoho CRM code and automation, and three years of changes. An AI developer working on it spent a lot of time guessing search terms and reading file after file.
I built a retrieval engine over all of it and plugged it into Claude Code through an MCP server, so a question returns the most relevant passages with their exact file and line.
- 82%
- of answers in the top 8, against 45% for keyword search
- 11,000+
- searchable chunks from 2,000+ files
- $0.56
- total cloud spend; the rest runs on a local GPU
What I built
- Indexing for seven kinds of source: documentation, C#, WordPress PHP, Deluge functions, CRM workflows, the CRM schema and git history
- Chunking by meaning: one C# method, one PHP hook or one CRM rule per chunk, with CRM settings rewritten as plain sentences
- Keyword and vector search run together, merged, then re-ranked by an AI model that reads each candidate against the question
- A benchmark of 136 real developer questions with known answers, used to test every change
- Re-indexing in seconds on every commit, with only git-tracked files indexed and every file scanned for secrets first
13,000 lines of C# and 271 automated tests, delivered in under two weeks.
Tools: C#, .NET 8, Roslyn, PostgreSQL, pgvector, Ollama, Qwen3 embedding and re-ranking models, Voyage AI, MCP, Claude Code
Diagrams and screenshots
