Building a RAG Pipeline for Semantic Code Search: A Developer Diary and Field Notes - The JetBrains Blog
Blog post from JetBrains
Building a production-grade RAG system for semantic code search requires more than embedding code chunks: it needs scalable vector storage, fast approximate nearest-neighbor search supplemented by exact reranking, and reliable metadata filtering for repositories, branches, folders, and access permissions. The authors describe using incremental, Merkle-tree-based indexing to reuse unchanged code across Git revisions, deduplicate embeddings, and avoid full reindexing, while separating vector ingestion from resource-intensive index construction and distributing customer data across shards and geographically located read replicas. Air Context also supports local on-device indexes for unpushed developer changes, merging local and remote results through reranking. After retrieval, the system removes duplicates, groups related code, ranks candidates with more careful models or methods, and formats results with detail proportional to relevance to conserve agent context. For more complex queries, a read-only retrieval sub-agent can iteratively search and inspect files within a controlled budget, allowing the main coding agent to focus on its assigned task; the authors report that agents using Air Context required up to 70% fewer turns in evaluations, with further assessment planned in the next installment.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| RAG | 7 | 350 | 93 | 53 | -60% |
| Vector Search | 5 | 1,097 | 201 | 87 | -45% |
| LLM | 2 | 3,355 | 742 | 191 | -49% |
| AI Agents | 1 | 5,564 | 1,101 | 236 | -21% |
| AI Coding Assistant | 1 | 1,698 | 401 | 139 | -13% |
| Agent Context Engineering | 1 | 109 | 47 | 35 | -55% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.