Answers a question about the caller org’s code with a CITED answer: retrieval packs grounding context, then the synthesizer writes the answer over exactly those spans, which come back alongside it. It never answers without grounding — with no matched code the answer is empty and says so, and with no synthesizer available the citations still come back with "degraded": true so the caller can reason over the spans itself.
Returns the INDEXED content of one file — read_file over the chunks the search tiers hold, for pulling up code an agent just found. It is NOT byte-verbatim: the git object plane is the source of record for exact bytes, history and blame. A file absent from the index is a 404, so an agent can tell "not indexed" from "empty file".
Finds code in the caller org’s index across three orthogonal retrieval tiers fused by reciprocal-rank fusion: lexical (FTS5 trigram over code-tokenized text), symbolic (real definition and reference edges), and semantic (embedding cosine over AST-boundary chunks). Pick one tier with type, or leave it to run all three as hybrid, which is what a coding agent usually wants. It is FAIL-HONEST: a retrieval outage answers 200 with an empty result set and "degraded": true rather than a 5xx, so an agent degrades instead of stalling. A malformed regex is a 400.
Returns one repository’s file structure with a per-file symbol count — get_repo_structure over the org’s own index, with no git checkout involved. A repository that has not been indexed answers an empty tree rather than an error, so an agent can tell "nothing here" without handling a failure.
Is askGet with the question in the request BODY, for a question too long or too awkward to put in a URL. query and repo in the body take precedence over ?q= and ?repo=; either source works alone.
Packs the most relevant code for a query into a token budget — THE primitive for a coding agent that has to decide what to put in a prompt. It retrieves seed spans, expands each with the definitions it calls and its key callers, then greedily fills the budget, so the answer is a coherent slice of the codebase rather than a list of disconnected matches. The top match is always included, truncated if it alone overflows, so a matched query never comes back empty. A retrieval outage answers 200 with an empty bundle rather than a 5xx.
(re)indexes a repository for the caller’s org, incrementally: files whose content hash is unchanged are skipped, so re-sending a whole tree is cheap. Each file is parsed for symbols, split at AST boundaries and — when the semantic tier is available — embedded, which is what makes it searchable across all three retrieval tiers. Pass prune to also DELETE indexed files absent from the request, which turns the call into a full sync; without it the call is an upsert. The index is written to the caller org’s own physically separate database.