Crate laminar_ai 

Backend-agnostic AI inference for LaminarDB.

AI SQL functions (ai_classify, ai_embed, ai_complete, …) are task contracts with fixed input/output. A named model from the ModelRegistry satisfies a contract; the model’s backend — local ONNX or a remote LLM — is hidden behind a single InferenceProvider. This crate holds the pieces that know nothing about Ring 0: the registry, the provider trait, request and response types, and (in later batches) the result cache, call log, adapters, and concrete backends.

Nothing here runs on the hot path. Inference is driven from a Ring 1 worker off the compute thread; see the operator in laminar-db.

Re-exports

pub use adapter::parse_response;

pub use adapter::AdapterError;

pub use cache::content_hash;

pub use cache::params_version;

pub use cache::AiCacheKey;

pub use cache::AiResultCache;

pub use cache::AiResultCacheConfig;

pub use cache::CachedOutput;

pub use call_log::AiCallLog;

pub use call_log::AiCallRecord;

pub use call_log::CallOutcome;

pub use provider::InferenceOutputs;

pub use provider::InferenceParams;

pub use provider::InferenceProvider;

pub use provider::InferenceRequest;

pub use provider::InferenceResponse;

pub use provider::ProviderError;

pub use provider::Usage;

pub use registry::BackendKind;

pub use registry::ModelBackend;

pub use registry::ModelEntry;

pub use registry::ModelRegistry;

pub use registry::RegistryError;

pub use registry::Task;

pub use runtime::AiRuntime;

pub use runtime::AiRuntimeError;

pub use runtime::ResolvedModel;

Modules

adapter

Resolve a raw provider response into a task’s per-row output, given the (task, backend) pair. Local classification takes argmax over logits (equal to argmax of softmax — classify returns only the label, so no softmax). Sentiment is numeric: local softmaxes the logits to P(pos) − P(neg), remote parses a number from the reply — a continuous score in [-1, 1]. Stays Arrow-free: returns InferenceOutputs; the operator builds the column.

backends

Concrete inference backends.

cache

Result cache for AI inference, keyed (content_hash, model_id, params_version).

call_log

Bounded in-memory log of inference calls, surfaced as laminar.ai_calls.

provider

The single transport abstraction over a model backend.

registry

The model registry: the curated catalog that maps a SQL-referenced model name to the backend that runs it and the tasks it can serve.

runtime

The assembled AI subsystem: the model registry, the provider clients that back it, the shared result cache, and the call log.