ingraph
A Rust framework for building retrieval systems — the kind that ingests media or documents, derives things from them (transcripts, embeddings, thumbnails, full-text documents), stores each derived thing in whatever engine suits it, and answers questions over the whole assembly.
Its one claim, stated before anything else:
cargo checkgreen means the retrieval assembly is connected.
Not compiles. Connected — every hop from the bytes that arrive to the answers that leave is a trait bound the compiler discharged. A vector index pointing at a record that does not exist, a question pipeline whose second step cannot eat the first step's output, an embedding of the wrong width — each is a compile error at the word the author wrote, not a runtime surprise at 3am.
0.1 is a first cut
Everything here may change or disappear in a later 0.x. No API is settled, no migration is
promised, and there is no deprecation cycle to rely on.
The surface is not a design that was drawn and then filled in. It is grown demand-driven: an item enters when something live asks for it — the codegen emits a path that has to resolve, a test names a type it cannot compile without, a downstream breaks for the want of it — and nothing is added because it will probably be needed later. That rule is what 0.1 was grown by, and it is what the next version will be grown by too.
Installation
[]
= "0.1"
One example
use Uuid;
Media is a root record: no lineage, so its key is its id alone. Keyframe declares its
lineage — the path from the root, root-first — so key = lineage ⊕ id spells
KeyframeKey { media_id, id } with nothing further said about it.
That expands into MediaKey, KeyframeKey, KeyframeLineage, an accessor family on each,
From<Record> and From<&Record> conversions, impl Transformer and impl Identifiable on the
record and on the key both, the per-database row and comparison lowering, the order-by vocabulary,
the keyset cursor, and the record's query unit.
Features
The crate is cut into tiers, and a build takes the rung it needs:
| feature | what it adds |
|---|---|
derive (default) |
the words a declaration is written in — implies runtime and flags |
runtime (default) |
the drive layer: nodes, channels, schedules, resources — implies engine |
engine |
the retrieve layer: gather, rank, fold, fusion, search towers |
| (none) | the vocabulary core alone — the value types, the filters, the comparison crossing |
On top of a tier sit the opt-in words, all default-off: the vocabularies a declaration may name
a field type from (smol_str, jiff, time, half, rust_decimal, regex, mediatime,
mediaframe), the backends a record is stored in (sqlite, postgres, mysql, mongo,
mongodb, tantivy), the wire face (async-graphql, dataloader), and the observation and
executor words (tracing, rayon, tokio-signal).
Platform support
Linux and macOS. Both are built, linted and tested in CI.
Windows is untested rather than broken. The library's own sources contain nothing Unix-specific,
so it may well compile there, but the repository cannot make that a claim it verifies: the intake
acceptance is driven by the tributaries filesystem watcher,
whose OS backends are FSEvents and inotify/fanotify only, so cargo test has no Windows
configuration to run. A Windows lane would assert support that no test covers.
Related crates
ingraph-derive— the macros, re-exported at this crate's root behind the defaultderivefeature. Nameingraphalone; this is here for the rare consumer that reaches the attributes directly.ingraph-codegen— the emission engine behind the macros. An implementation detail ofingraph-derive, published because it is a dependency of one.
License
ingraph is under the terms of both the MIT license and the Apache License (Version 2.0).
See LICENSE-APACHE and LICENSE-MIT for details.