Skip to main content

​
1. File discovery

unch index walks the repository, applies .gitignore plus explicit --exclude patterns, and decides which source files should be processed.

​
2. File hashing

Each candidate file is hashed. The active file-hash state in .semsearch/filehashes.db is used to decide whether the file can be reused or needs to be reindexed.

​
3. Symbol extraction

For supported languages, unch uses Tree-sitter to extract top-level symbols and attached docs. That currently covers:
  • Go
  • Rust
  • TypeScript
  • JavaScript
  • Python
For unsupported files or parser failures, index can still use the legacy prefix fallback.

​
4. Embedding generation

Each extracted symbol is flattened into an indexed document and embedded with the selected provider. Provider options today:
  • llama.cpp for local GGUF models through yzma
  • openrouter for remote embedding APIs
Known model ids today:
  • embeddinggemma
  • qwen3

​
5. Snapshot activation

Embeddings and symbols are written into provider-scoped and model-scoped snapshots in .semsearch/index.db. When indexing succeeds, the new snapshot becomes active for that provider/model pair. Other provider/model snapshots are left untouched. Embedding vector tables are also separated by dimension, so models with different embedding sizes can coexist in one .semsearch directory. unch search supports:
  • semantic
  • lexical
  • auto
auto stays semantic-first, but can prefer lexical results when the query looks more like a symbol name or code fragment.

​
7. Optional remote restore

If the manifest is bound to remote CI, unch search can restore the latest compatible published state before executing the query.