Using the JetBrains program structure interface for codebase context (blog.sweep.dev)

Sweep announced a JetBrains-focused autocomplete that uses the IDE’s Program Structure Interface (PSI) to provide precise, low-latency codebase context to its in-house LLM-powered completion engine. After early feedback that their autocomplete hallucinated methods (e.g., suggesting nonexistent DatabaseClient methods), Sweep rejected sending whole repos (context length hurts model accuracy and adds latency) and common search approaches. TF-IDF can’t identify which token is the real definition vs. noise, and vector search incurs privacy, indexing, and latency costs (client indexes are heavy; server indexes expose code). Frontier models also degrade on long contexts and each ~10k tokens adds ~100ms latency—unacceptable inside their 100ms budget. Technically, because JetBrains PSI runs in-process and maintains an up-to-date, incremental AST of the whole project, Sweep’s plugin performs direct lookups of definitions around the cursor (30ms cold, <1ms after cache). Combined with their self-hosted inference stack (speculative decoding optimizations and regional GPUs to cut round-trip time), this yields accurate, definition-aware suggestions without embeddings or document search. The result: a 3% lift in autocomplete acceptance with negligible added latency. Sweep’s plugin is available in the JetBrains marketplace and the team is soliciting early feedback via Discord.