Language coverage

prom.codes does not chunk source by line windows. Every supported language is parsed with Tree-sitter into a real syntax tree, chunked along symbol boundaries (function, class, method, interface), and linked through a reference graph. That is why a query lands on a symbol, not a paragraph that happens to mention the right word.

Supported languages

The indexer ships 15 language IDs today. Detection is by file extension; unknown code extensions are skipped, not guessed. Documentation and curated config files are handled separately — see Beyond code: prose & config below.

Vue is the one composite format. There is no maintained Vue Tree-sitter grammar we can ship, so an .vue single-file component is split into its blocks instead: the HTML grammar gives the layout, and each <script> body is re-parsed with the TypeScript or JavaScript grammar its lang attribute names. You get the script's own symbols, one component symbol per file, one property per declared prop and emit, and a reference for every component tag used in the template — so find_callers(UserCard) answers "who renders this?". <style> blocks are not indexed.

Each language has both a symbol extractor (definitions, exports, properties) and a reference extractor (imports, calls, member access). The reference graph is what powers find_callers, find_callees, and the 1-hop context expansion around a hit.

Measured retrieval quality

We do not claim parity by assertion. Every non-trivial language is validated with the same self-index engine-lift harness — index a real OSS repo, synthesise queries from its own doc comments, and measure ranking quality against held-out gold targets with our production retrieval stack.

The headline metric is NDCG@10 for the pre-graph ranking (the surface an agent sees before graph expansion). Higher is better; 1.0 is a perfect ranking.

Why the graph matters

Switching from hybrid-no-graph to hybrid-full trades a little top-of-list precision for a real jump in recall, consistently across every language we measured. For Ruby, Recall@100 climbs from 0.900 to 0.976; for Kotlin, from 0.774 to 0.874. That recall lift only appears if the reference extractor joins calls to their definitions correctly — so the lift doubles as a correctness signal for each new language we add.

Beyond code: prose & config

A codebase is more than its source. READMEs, ADRs, specs and a handful of key config files (package.json, tsconfig.json, docker-compose.yml, pyproject.toml …) answer questions an agent asks constantly — "how is this configured?", "where is this documented?". Tree-sitter does not help here, so these files travel a separate grammar-less document path: pure-string extractors, no parser.

The hard part is not indexing everything. Naively swallowing every .json — lockfiles, fixtures, generated blobs — floods the corpus with distractors and measurably lowers retrieval quality for the actual code. So the document path is deliberately selective, in three tiers:

Two outcomes follow from this. A non-code file always yields one document unit, so it is retrievable as a whole even when it has no internal structure; richer files additionally yield section units (a Markdown heading block, a config key or table). And because config is allowlist-gated rather than extension-gated, the high-signal files land in the corpus while the noise stays out — the code numbers above are protected, not diluted.

Not yet supported

Swift is deliberately deferred. Its current Tree-sitter grammar targets a newer ABI (15) than our pinned runtime, so wiring it in safely needs a tree-sitter@0.22+ upgrade first. We would rather ship it on a stable runtime than rush a fragile binding.

If you need a language that is not on the list, the per-language cost is small (a symbol extractor and a reference extractor that mirror the existing ones) — tell us which one and we will benchmark it.