Retrieval lab
See what retrieval actually does.
RAG, without the fog
Retrieval-augmented generation is a process, not a magic database.
A question first retrieves relevant evidence. That evidence augments the prompt. A model may then generate an answer from the supplied context. Raccoon currently exposes the first two steps so you can inspect them without paying for inference.
01 · Corpus
Connect a real repository
What is a chunk? A repository is too large to hand to a model whole. Raccoon divides files and historical records into smaller evidence units that can be searched, ranked, and cited independently.
Turn the question into searchable terms
Raccoon keeps exact identifiers intact, then splits ordinary language into normalized terms. This matters because a symbol like SyncCoordinator carries more signal than a generic word like system.
Today: lowercase terms, camelCase splitting, and identifier-friendly tokenization.02 · Configuration
Retrieval controls
What this knob means
Looks for shared words, identifiers, and paths. Rare matches count more than common ones, which makes lexical search excellent for function names, filenames, errors, and issue numbers.
Turn this off to remove exact-term retrieval. It is required right now because dense search is not connected yet.What this knob means
Dense retrieval converts the query and every chunk into vectors. Nearby vectors can match similar meaning even when the words differ, such as login and authentication.
Once connected, this knob will let you compare semantic retrieval against lexical retrieval and fuse both ranked lists.What this knob means
Current source explains how the code works now. Commits, issues, and pull requests can explain how it changed and why a decision was made.
History is eligible, so rationale may rank beside current code.Run a query to inspect the retrieval path.
03 · Evidence
Retrieved context
The dumpster is empty.
Connect a repository to index its current code, documentation, and recent history.