rig-model-catalog

Provider-agnostic model metadata for the Rig ecosystem.

rig-model-catalog answers one question consistently across providers:

Given a model id, what is its context window, what capabilities does it advertise, and what do I need to know to size prompts and budget tokens?

It ships a single trait — ModelMetaProbe — plus a plain-data ModelDescriptor and a handful of backends (live Ollama, static catalogs for OpenAI / Anthropic, and a stub for tests). The default build pulls no HTTP client or provider SDKs; backends are opt-in via Cargo features.

Status

v0.1.0 — initial release. API is #[non_exhaustive] where it counts; additive changes (new descriptor fields, new probes, new capability variants) are not breaking.

Quick start

[dependencies]
rig-model-catalog = { version = "0.1", features = ["ollama", "static"] }

use rig_model_catalog::{ChainedProbe, ModelMetaProbe, OllamaProbe, StaticProbe};

# async fn run() -> anyhow::Result<()> {
let probe = ChainedProbe::new(
    OllamaProbe::live("http://localhost:11434"),
    StaticProbe::builtin(),
);

if let Some(desc) = probe.describe("gpt-4o").await? {
    println!("context window: {:?}", desc.context_window);
    println!("capabilities:   {:?}", desc.capabilities);
}
# Ok(()) }

Thinking-capable models

Capability::Thinking means the model advertises a native reasoning channel. For Ollama models used through Rig, callers should opt into that native separation with think: true and still assert that user-visible output is free of raw <think> / </think> tags.

use rig_model_catalog::{Capability, ModelMetaProbe, OllamaProbe};

let base_url = "http://localhost:11434";
let model_name = "qwen3.5:9b";
let model = ollama_client.completion_model(model_name);

let probe = OllamaProbe::live(base_url);
let descriptor = probe.describe(model_name).await?;
let supports_thinking = descriptor
  .as_ref()
  .is_some_and(|d| d.capabilities.contains(&Capability::Thinking));

let mut builder = rig::agent::AgentBuilder::new(model);
if supports_thinking {
  builder = builder.additional_params(serde_json::json!({ "think": true }));
}
let agent = builder.build();

The capability is not a sanitizer. It tells the host which provider contract to activate. Applications that render model output directly should keep a live smoke check equivalent to !output.contains("<think>") && !output.contains("</think>") so provider or model regressions are caught before users see reasoning tags.

Features

Feature	Pulls	What you get
default	—	ModelMetaProbe, ModelDescriptor, StubProbe, ChainedProbe
ollama	reqwest	OllamaProbe::live(...) against POST /api/show
static	—	StaticProbe::builtin() with bundled OpenAI + Anthropic tables
rig-hook	rig-core	MetaHook + HookPair (PromptHook telemetry composition)
observe	rig-hook	Re-emits MetaHook prompt lifecycle events on the rig_tap target
pricing	—	PricingTable + ModelPrice with bundled OpenAI + Anthropic USD/M rates and snapshot metadata

Design rules

• Unknown is normal. Probes return Ok(None) for models they don't recognise. Err is reserved for hard failures (transport down, auth rejection, malformed response).
• No tokio in [dependencies]. The library is runtime-agnostic; tests and examples use tokio via dev-deps.
• No panics in library code. Every fallible operation is a Result or an Option. Clippy unwrap_used, expect_used, panic, indexing_slicing, todo, unimplemented, unreachable, dbg_macro, and await_holding_lock are deny/forbid.
• #[non_exhaustive] on the public payload. Adding fields to ModelDescriptor or variants to Capability is not a breaking change.

Roadmap

Released in v0.1.0. Unreleased on main and exercised in the CI matrix: MetaHook + HookPair (rig-hook feature), MetaHook rig_tap emission (observe feature), PricingTable + ModelPrice (pricing feature), RuntimeDescriptor / OllamaProbe::runtime, ModelMetaProbeDyn + DynProbe, and a TTL Cache<P> memoiser. See CHANGELOG.md for the full list.

PricingTable::builtin() exposes the bundled pricing snapshot date and provenance through snapshot_date() and snapshot_provenance() so downstream cost reports can warn on stale static rates while still overriding prices with host-owned JSON or programmatic entries.

Still planned:

• Knowledge-cutoff + deprecation fields (additive, data files).
• Bedrock and llama.cpp probes.
• CLI: rig-model-catalog show <provider>:<model>.

License

Dual-licensed under MIT or Apache-2.0 at your option.