RustUse/use-ai

Composable AI interaction primitives for RustUse.

use-ai is a focused RustUse set for representing prompts, messages, context windows, roles, model and provider metadata, capabilities, tool calls, agents, reasoning, planning, RAG metadata, memory metadata, guardrails, and AI-specific evaluation as small Rust primitives. It is not an AI runtime, model client, agent orchestrator, vector database, memory store, guardrail engine, or eval runner.

Workspace crates

Crate	Path	Purpose
use-ai	crates/use-ai/	Facade over the focused AI interaction primitive crates
use-ai-prompt	crates/use-ai-prompt/	Prompt names, text, variables, formats, parts, and status
use-ai-message	crates/use-ai-message/	Message identity, roles, parts, formats, token counts, state
use-ai-context	crates/use-ai-context/	Context items, citations, priorities, fit, and windows
use-ai-role	crates/use-ai-role/	Participants, authority, role scopes, personas, and status
use-ai-model	crates/use-ai-model/	AI model identity, modalities, interfaces, limits, lifecycle
use-ai-provider	crates/use-ai-provider/	Provider, endpoint, quota, billing, API mode, and region labels
use-ai-capability	crates/use-ai-capability/	Capability support labels for AI model and provider surfaces
use-tool-call	crates/use-tool-call/	Tool-call names, arguments, result kinds, schemas, and status
use-agent	crates/use-agent/	Agent identity, modes, loops, actions, observations, handoffs
use-reasoning	crates/use-reasoning/	Reasoning modes, visibility, step kinds, effort, and artifacts
use-planning	crates/use-planning/	Plan identity, steps, status, dependencies, review, and risks
use-rag	crates/use-rag/	RAG corpus, document, chunk, retrieval, ranking, and grounding
use-ai-memory	crates/use-ai-memory/	Memory identity, confidence, scope, retention, sensitivity
use-guardrail	crates/use-guardrail/	Guardrail identity, checks, actions, results, and violations
use-ai-eval	crates/use-ai-eval/	AI eval runs, rubrics, scores, datasets, metrics, and outcomes

Experimental

Every crate in this workspace is experimental while the release line remains below 0.3.0. Expect small API adjustments as the AI primitive surface settles.

Relationship to use-ml

use-ai models AI interaction primitives: prompts, messages, roles, context windows, tool calls, agents, RAG, reasoning, memory, guardrails, AI model interfaces, and AI-specific evaluation.

use-ml models machine-learning primitives: datasets, features, labels, tensors, model artifacts, training, inference, evaluation, metrics, pipelines, embeddings, experiments, and model documentation.

These sets are siblings. They should interoperate conceptually but avoid dependency cycles.

Example

use use_ai::{AiMessageRole, AiModelName, PromptFormat, PromptName, ToolName};

let prompt = PromptName::new("support-triage")?;
let model = AiModelName::new("reasoning-chat")?;
let tool = ToolName::new("ticket-search")?;

assert_eq!(prompt.as_str(), "support-triage");
assert_eq!(model.value(), "reasoning-chat");
assert_eq!(tool.as_str(), "ticket-search");
assert_eq!(AiMessageRole::Assistant.as_str(), "assistant");
assert_eq!("plain text".parse::<PromptFormat>()?, PromptFormat::PlainText);
# Ok::<(), Box<dyn std::error::Error>>(())

Scope

• Validated identifiers, names, text references, finite scores, bounded counts, and positive limits.
• Small enums for prompts, messages, context, roles, model metadata, providers, capabilities, tools, agents, reasoning, planning, RAG, memory, guardrails, and AI evaluation.
• Metadata-only primitives suitable for application glue code, docs tooling, examples, CLIs, and test fixtures.

v0.2 follow-ups

Keep future use-ai work focused on AI interaction metadata. Good follow-up candidates include optional serde support, richer schema descriptors, provider-neutral request metadata, optional no_std compatibility where practical, and conversion helpers between sibling crates where they do not create dependency cycles.

Non-goals

• Calling models, providers, local runtimes, model hubs, cloud services, or APIs.
• Running tools, shell commands, retrieval, embeddings, guardrails, memory storage, evals, planners, or agent loops.
• Training models, serving models, implementing tensor math, allocating vectors, or computing embedding values.
• Shelling out to Python, Node.js, model CLIs, cloud CLIs, vector databases, notebooks, package registries, or external services.

Development

cargo fmt
cargo clippy --workspace --all-targets --all-features
cargo test --workspace --all-features

For release-readiness checks, also run the repository tasks in the Makefile.

License

Licensed under either of the following, at your option:

• Apache License, Version 2.0
• MIT license