aire

A Python library for composing models, retrieval, tools, agents, and workflows. One AI facade exposes the components; lower-level classes let you replace them individually. Apache-2.0 · Python 3.11+ · distribution: aire-ai · import: aire.

Alpha, 0.4.x. Start with the offline defaults, test your chosen adapters, and read the limitations before using it for production workloads. The echo model tests plumbing; it does not reason or answer questions intelligently.

from aire import AI

assistant = (
    AI.project("handbook")
    .documents(["Access requires a short lived token and an authorized role."])
    .model("mock:echo")
    .embedder("local:hashing")
    .citations(True)
)
report = assistant.index()
answer = assistant.ask("How does access work?")
print(report.documents, report.chunks)
print(answer.text)       # Echoes the assembled retrieval prompt, including context.
print(answer.citations)  # Structured references to retrieved chunks.

This example uses no credentials, model downloads, or network services. With the package installed, it runs as written. The same pipeline can use a hosted model or a local inference server by changing the model reference.

Install and develop

python -m pip install aire-ai

The PyPI name aire belongs to a different project. Install aire-ai. To work on this source tree, including the reliability changes described below:

git clone https://github.com/desenyon/aire.git
cd aire
python3.13 -m venv .venv
source .venv/bin/activate             # Windows: .venv\Scripts\activate
python -m pip install -e '.[dev,docs]' build
python examples/reliable_local/main.py

The last command is an offline integration smoke test. It exercises SQLite persistence, document replacement, retrieval with citations, a scripted tool call, an agent session, a file-queue job, and workflow resume in a temporary directory that is cleaned up afterward.

Provider-named extras such as openai, anthropic, and ollama add no vendor SDK: these adapters use the core httpx dependency. all installs many heavy optional packages; it is unnecessary for agents, local RAG, or the default tests. The complete list is in pyproject.toml.

Architecture

flowchart TD
    APP[Python application or CLI] --> AI[AI facade / Assistant builder]
    AI --> RT[Runtime: settings, registries, events, resources]
    AI --> MODEL[Model / EmbeddingModel]
    MODEL --> HTTP[Provider HTTP adapters or offline builtins]
    AI --> KB[Knowledge pipeline]
    KB --> PREP[Load, chunk, embed, validate]
    PREP --> STORE[Local / SQLite / service vector stores]
    STORE --> RET[Retrieve, rerank, compress, cite]
    RET --> MODEL
    AI --> AGENT[AgentExecutor: budgeted model and tool loop]
    AGENT --> MODEL
    AGENT --> TOOLS[Tool schemas, policy, approval, execution]
    AI --> WF[Workflow graph and checkpoints]
    WF --> JOBS[In-process / file / Redis workers]
    RT --> SUPPORT[Safety, tracing, evaluation, lifecycle]

There are three useful entry levels:

1. Project builder: AI.project(...).documents(...).model(...), followed by index(), ask(), evaluate(), or deploy().
2. Composable facade: AI.models, AI.rag, AI.agents, AI.workflows, AI.workers, AI.gateway, and the other namespaces exposed by AI.describe().
3. Direct components: instantiate Runtime, Knowledge, Agent, Workflow, or provider/store classes when you need explicit dependency ownership.

Runtime owns settings, registries, provider plugins, event dispatch, and registered HTTP resources. The facade creates a process-wide runtime lazily. Use an explicitly owned runtime for isolation, and close it with await runtime.aclose() or async with Runtime(...). Close directly created SQLite stores with await store.aclose().

The architecture guide and public API map provide more detail. These are composable systems: configuring a workflow does not automatically deploy it, attach tracing, or run evaluations.

Models and streaming

import asyncio
from aire import AI
from aire.models.types import GenerationRequest

async def main():
    model = await AI.models.use("mock:echo")
    print(await model.ask("Hello, aire"))
    async for chunk in model.stream(GenerationRequest.of("Stream this text")):
        print(chunk.text, end="")
    await AI.runtime().aclose()

asyncio.run(main())

Use await APIs inside notebooks, web handlers, and existing event loops. *_sync methods and aire.models.base.run_sync() are conveniences for synchronous programs; they cannot be called inside an active event loop. The low-level model contract is generate(GenerationRequest), stream(...), info, and health().

References have the form provider:name. Builtins include mock:echo, local:hashing for embeddings, and registered callable:<name> models. HTTP providers include OpenAI, Anthropic, Ollama, Hugging Face and OpenAI-compatible aliases. Provider capabilities differ; a shared interface does not imply every model supports tools, structured output, media, or identical token accounting.

from aire import AI

# Select a model available to your provider account / local server.
# These calls require the corresponding server or credentials when used.
model = AI.models.use_sync("openai:gpt-4o-mini")
local_model = AI.models.use_sync("ollama:llama3.2")

Message.tool_calls preserves assistant declarations alongside tool-result messages and their tool_call_id. The existing import from aire.models.types import ToolCall remains valid. GenerationResult.message produces a complete assistant message, suitable for the next model request.

OpenAI-compatible streams buffer argument fragments by call index and emit each complete call once at the finish event. Anthropic streams assemble tool blocks and emit them when the block closes. Both preserve usage and finish reasons; malformed or unfinished tool arguments raise structured ProviderError rather than becoming an executable partial call. Ollama emits complete tool calls from its JSON stream. Models without native streaming use a one-chunk fallback that preserves text, calls, usage, and finish reason.

Exact response caches retain tool calls during capture and replay. The model semantic-cache wrapper bypasses requests with tool declarations or tool history; similar text is insufficient to identify equivalent tool arguments.

Retrieval and index updates

The Knowledge pipeline loads documents, chunks their text, computes embeddings, and publishes chunks to a store. Queries combine vector similarity and lexical search by default, then rerank and optionally compress context before generation. Answer exposes the generated text, citations, usage, and retrieval metadata. Citations identify retrieved evidence; they do not independently prove the answer.

import asyncio
from aire.core.config import Settings
from aire.core.runtime import Runtime
from aire.models.builtin import HashingEmbedder
from aire.rag.pipeline import Knowledge
from aire.rag.sqlite import SQLiteVectorStore
from aire.rag.types import Document

async def main():
    async with Runtime(Settings()) as runtime:
        store = SQLiteVectorStore("knowledge.sqlite")
        try:
            knowledge = Knowledge(runtime, store=store, embedder=HashingEmbedder())
            await knowledge.ingest([
                Document(id="access-guide", text="Access requires a token.",
                         metadata={"tenant": "demo", "source": "handbook"})
            ])
            await knowledge.reindex_document(
                Document(id="access-guide", text="Access requires a token and role checks.",
                         metadata={"tenant": "demo", "source": "handbook"})
            )
            hits = await knowledge.retrieve("access", filter={"tenant": "demo"})
            print([hit.chunk.text for hit in hits])
        finally:
            await store.aclose()

asyncio.run(main())

Use stable Document.id values for updates. Repeated ingest() calls append new chunks; they are not automatic deduplication. reindex_document() replaces all chunks belonging to that document ID. Empty/invalid documents or invalid embedding batches fail before publication, preserving the existing index.

IncrementalIndex(knowledge).reindex(source, clear=True) (or knowledge.ingest(source, replace=True)) prepares a full replacement before publishing. Local and SQLite support atomic full replacement. Other adapters must override replace_all(); otherwise they raise rag.atomic_replace_unsupported without clearing the store. clear=False keeps append/upsert behavior.

Local/SQLite filters apply before the result limit, including empty and punctuation-only queries. Plain filter keys use exact metadata equality; {"__acl__": {"tenant": "demo", "roles": ["reader"]}} enables ACL matching. k <= 0 returns no local/SQLite hits. See RAG for query rewriting, reranking, compression and remote-store behavior. Treat ACL metadata as trusted application data and verify each remote backend's filtering semantics.

Agents and tools

Tools expose a typed function, a JSON schema, permissions, and a side-effect classification. The agent loop asks a model for text or calls, applies policy and approval checks, executes allowed tools, and sends observations back to the model. A tool being registered does not mean an echo model will choose it. This deterministic example scripts one call explicitly:

from aire import AI
from aire.models.builtin import EchoModel
from aire.models.types import ToolCall

@AI.tool
def add(a: int, b: int) -> int:
    """Add two integers."""
    return a + b

model = EchoModel()
model.scripted_tool_calls = [ToolCall(id="addition", name="add", arguments={"a": 2, "b": 3})]
agent = AI.agents.create_sync(model, tools=[add], session="agent-session.json")
result = agent.run_sync("Use the addition tool.")
print(result.status)
print([m.text_content for m in agent.state.messages if m.role == "tool"])  # ['5']

Configure aire.agents.types.AgentConfig for step/token/cost budgets and tool parallelism. Approval-dependent actions are denied unless an approver allows them. Inspect AgentResult.steps, status, error, and usage; an output string alone does not establish success. The final echo answer remains an echo, even when the scripted tool has executed correctly.

Memory retains assistant declarations and tool observations together; a truncated buffer drops orphan tool-result messages at its beginning. JSON sessions preserve the same typed messages. Session saves are atomic snapshots, but a session is not a transaction log of every side effect: normal persistence happens at turn end or explicit pause. See agents for planning, teams, memory, streamed step events, and approvals.

Workflows and recovery

import asyncio
from aire import AI

async def main():
    flow = AI.workflow("report", checkpoint_path="report-state.json", max_visits=3)
    flow.add("prepare", lambda value, ctx: value.strip())
    flow.add("render", lambda value, ctx: f"Report: {value}")
    flow.connect("prepare", "render")
    result = await flow.run("  hello  ")
    assert result.ok and result.output == "Report: hello"
    resumed = await flow.resume()  # Finished work is not repeated.
    assert resumed.output == result.output

asyncio.run(main())

The first added node is the entry unless entry(name) overrides it. Nodes receive (input, context) and may return a value or an awaitable. A single completed predecessor supplies its output; multiple completed predecessors supply a {name: output} mapping. This input rule includes completed predecessors whose conditional edge did not fire. An edge condition controls scheduling.

A node is ready when at least one incoming edge fires and all predecessors are terminal. Untaken branches propagate SKIPPED to a fixed point, allowing joins to proceed regardless of node registration order. Ready nodes run in concurrent waves. Synchronous functions run on the event-loop thread; only awaitable node bodies can be timed out or cooperatively cancelled. Use async nodes for I/O.

Checkpoints now include pending/in-flight nodes, edge firing/consumption counts, and node visit counts. Failed nodes may resume within their remaining visit budget; exhausted failures stay failures. Cycles are bounded by max_visits. Closing a streamed run cancels unfinished tasks. An unresolved/disconnected graph reports a failure rather than silently completing.

Recreate the same graph and functions before resuming a saved state. Checkpoints store data, not executable code. Nodes completed in memory but not checkpointed may run again after interruption. Make side effects idempotent, and use one checkpoint path per execution. Older DAG checkpoints remain readable; older cyclic checkpoints lack enough history to reconstruct exact edge counts.

Job workers

import asyncio
from aire import AI

async def main():
    flow = AI.workflow("double").add("double", lambda value, ctx: value * 2)
    worker = AI.workers.create("file", directory="jobs", workflows={"double": flow})
    job = worker.enqueue("double", 21)
    result = (await worker.drain(max_jobs=10))[0]
    assert result.job.id == job.id and result.job.result == 42

asyncio.run(main())

Queue execution preserves the original job ID and creation time. Unknown workflows become retained failed results. Cancellation, invalid job JSON, and result-write failures leave the claim available for diagnosis and recovery. Malformed jobs need repair or removal before replay.

After stopping all workers (including Redis blocking calls), call worker.recover() to requeue unacknowledged claims. The file worker also cleans up claims with an already-saved result. Recovery is explicit: there is no lease, automatic watchdog, dead-letter retry policy, or exactly-once execution. A crash between a side effect and acknowledgement can replay the side effect. Redis results accumulate until the operator removes them; the implementation assumes one Redis instance/keyspace and does not manage Redis Cluster hash slots.

Configuration

Settings.load() searches the current directory and its parents for aire.yaml, aire.yml, or aire.json, or accepts an explicit path. Current precedence is:

1. Explicit Python overrides.
2. Environment variables such as AIRE_MODEL__REF (double underscore nests keys).
3. The selected project file.
4. ~/.config/aire/config.yaml, used only when no project file or explicit path is selected.
5. Library defaults.

pyproject.toml contains package/tool configuration; Settings.load() does not read a [tool.aire] table. AI.from_config(path) configures a project runtime; attach document sources with .documents(...) in Python.

# aire.yaml
project: handbook
model:
  ref: mock:echo
  embedder: local:hashing
agent:
  max_steps: 12
observability:
  enabled: true
  exporter: memory

Call AI.configure() to load the defaults/environment, or AI.configure(model={"ref": "mock:echo"}) for an explicit override. Namespaced settings do not automatically configure every directly instantiated object; pass its documented constructor options (for example AgentConfig) explicitly.

Provider adapters accept explicit api_key/base_url options. Credentials may also come from providers.<name> configuration or standard environment variables such as OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENAI_BASE_URL, and OLLAMA_HOST. Keep credentials in your environment or secret manager, not committed examples. The adapter's available model list, pricing, and capabilities remain provider specific; built-in cost tables are estimates, not billing guarantees.

CLI, serving, evaluation and extensions

aire --help
aire doctor
aire run 'hello, aire'

assistant.deploy() constructs a FastAPI application; it does not publish a service. The gateway provides OpenAI-compatible and Anthropic-compatible routes, model aliases and routing. See gateway for authentication, configuration and route details. Deployment file generators produce starting points that need environment-specific review.

Use assistant.evaluate(...) or the evaluation namespace with explicit datasets and metrics. Some metrics are deterministic lexical checks; model/embedding judges require their configured dependencies. Offline wiring tests do not measure answer quality. Evaluation examples demonstrate gates. Tracing and metrics expose operational events when configured; they do not replace application monitoring, audit storage, or provider billing records.

MCP, OpenAPI tools, and custom model/store implementations provide extension points. Optional packages are imported lazily. Most components expose .describe() for capability inspection.

Validation

From the activated source environment:

make lint                         # Ruff lint and formatting
make typecheck                    # Strict mypy over src/aire
make test                         # Offline suite; live tests excluded
python examples/reliable_local/main.py
python -m mkdocs build --strict
python -m build                   # Wheel and source distribution

CI tests Python 3.11, 3.12 and 3.13, runs the offline smoke, and validates docs and packaging on Python 3.12. Core tests do not require PyTorch, Redis, paid providers, GPU training, or model downloads. Optional dependency tests verify missing-dependency errors where the backend is absent. tests/live/ is opt-in and requires explicitly configured services; passing offline tests does not verify those integrations.

Regression tests cover tool declaration round trips, interleaved streaming arguments, usage-only events, invalid/incomplete tool streams, metadata/ACL filtering, embedding failures, SQLite rollback, atomic snapshots, failed queue acknowledgements, recovery, skipped workflow branches, visit budgets, checkpoint compatibility and cancellation. Tests use synthetic data and injected failures.

Migration and reliability changes

The AI facade, offline defaults, provider:name references and existing provider constructors remain. See migration and CHANGELOG for the complete changes and operational details:

• Existing messages without tool_calls load with an empty list. Historical tool declarations that were never saved cannot be recovered automatically.
• Tool calls in streams are complete objects emitted once, not partial argument fragments. Consumers should execute only complete calls.
• Full-index replacement on adapters without an atomic override now fails before mutation. Implement replace_all() or use an explicit backend-managed swap.
• JSON/JSONL snapshot publication is atomic and temporary files are private. Multiple writers still use last-writer-wins semantics; this is not a database lock.
• Workflow checkpoints add scheduling fields; legacy DAG checkpoints still load. Use new checkpoints for exact cycle accounting and keep graph definitions stable.
• Stop old Redis workers before upgrading; the reliable queue adds processing and result keys. Stop all workers before manual recovery.

Limitations

• Alpha APIs may change. No production SLA or unmeasured performance/quality claim.
• mock:echo echoes text; local:hashing is deterministic lexical feature hashing, not a trained semantic embedder. Local tokenization primarily handles ASCII letters/digits; multilingual retrieval needs an appropriate embedding backend.
• Local/SQLite search is an in-memory scan. SQLite persistence does not turn it into an ANN index or keep independent processes' caches synchronized.
• Remote store replacement/filter guarantees vary by adapter. The generic document fallback enumerates at most one million hits and cannot promise a complete remote document listing. Remote write/delete failure can leave duplicate generations.
• File queues are local-only. Queue/workflow recovery is at-least-once, and atomic snapshots do not guarantee survival of every hardware/power failure.
• Session persistence is turn/pause based. It does not guarantee restart at every tool boundary or enforce a cumulative budget across separate process resumes.
• Guardrails and policies are application controls, not a sandbox for arbitrary Python code. Review tool permissions and side effects.
• Some vision/audio pipelines explicitly return stub=True. Foundation helpers build toy architectures unless pretrained loading is requested. Dry-run training is not evidence of a trained model. See honesty and GAPS.

Project resources

Capability master plan · Repository audit · Open gaps

Documentation · Runnable examples · Contributing · Security · Code of conduct · Citation · License

The 0.4 capability contracts map every public namespace to tests, supported backends, and explicit limits. See the release QA record for current validation.