Ragwell Python SDK

Typed synchronous and asynchronous Python clients for Ragwell, a managed service for document ingestion and retrieval with source citations.

Status: 0.2.0 developer beta release candidate. This source implements the reviewed machine API, including optional Jev reranking. The candidate runtime passed the deployed beta lifecycle and a bounded live Jev smoke; a fresh hosted CI matrix remains a release gate. The SDK returns retrieval evidence and citations; it does not generate answers.

Requirements and installation

Python 3.11 or newer is required. Install the ragwell distribution from PyPI:

python -m pip install ragwell

For local development:

uv sync --locked --python 3.11

Runtime use requires an existing project grant and API key. The SDK does not create projects, users, keys, or billing resources. Supply the endpoint and key explicitly or through RAGWELL_BASE_URL and RAGWELL_API_KEY; explicit arguments win. No beta hostname is built into the package and no .env file is loaded implicitly.

Synchronous quick start

import os
from pathlib import Path

from ragwell import Ragwell
from ragwell.types import SearchFilters

with Ragwell() as client:
    project = client.project(os.environ["RAGWELL_PROJECT_ID"])
    ready = project.documents.upload_and_wait(
        file=Path("guide.pdf"),
        wait_timeout=300,
    )
    results = project.search(
        query="What is the retention policy?",
        filters=SearchFilters(document_ids=[ready.intake.document.id]),
    )
    for item in results.items:
        print(item.rank, item.source_filename, item.parts)

Asynchronous quick start

import os

from ragwell import AsyncRagwell


async def retrieve(query: str):
    async with AsyncRagwell() as client:
        project = client.project(os.environ["RAGWELL_PROJECT_ID"])
        return await project.search(query=query)

Initialize one async client at application startup, share it within its owning event loop, and close it at shutdown. Injected HTTPX clients are borrowed unless owns_http_client=True; SDK-created pools are owned and closed by the SDK.

Public resource layout

client.project(project_id) is local and performs no preflight request. Both client styles expose the same 26 contract operations:

• client.upload_policy.get()
• project.get() and project.embedding_connection.get()
• project.uploads.create(), .get(), .upload_content(), and .finalize()
• project.documents.list(), .get(), .replace_metadata(), .delete(), .inspect(), and .activity()
• project.documents.generations.list(), .chunks.list(), and .sources.get()
• project.documents.exports.create(), .get(), .stream_part(), and .download_part()
• project.jobs.list(), .get(), .retry(), and .cancel()
• project.deletions.get() and .retry()
• project.search()

Convenience workflows add documents.upload(), documents.upload_and_wait(), jobs.wait(), deletions.wait(), documents.exports.wait(), and lazy .iter() methods for documents, jobs, generations, and chunks.

Generated public types and enums are available from ragwell.types. They preserve UUIDs, aware timestamps, omission through UNSET, explicit null, complete source parts and scores, command acceptance versus current resource state, and additive response fields through additional_properties.

Reliability and ownership

• HTTP 202 means accepted, never completed. Waiters have finite deadlines and never cancel or retry remote work implicitly.
• Job cursors are opaque and filter/project scoped. Iterators never parse or create them, stop at null, and reject repeated cursors.
• Safe reads receive at most two bounded retries. Replayed mutations keep one idempotency key for the logical command. Search is never automatically replayed.
• command_in_progress may retry within the deadline using Retry-After; quota, authorization, validation, and revision/idempotency conflicts do not.
• Upload helpers accept paths, bytes, and seekable binary streams, hash incrementally, restore caller cursors, and never close caller-owned streams. Bytes/streams need an explicit filename.
• Export saves stream to an SDK-owned temporary file in the destination directory, verifies manifest byte size and SHA-256, then replaces atomically. Overwrite is opt-in.

See deadlines, bounds and recovery for timeout-zero behavior, transfer budgets, cancellation, decoded-body caps and accepted-intake recovery IDs.

Errors derive from RagwellError. ApiError preserves safe status/code, field errors, quota detail, retry hints, operation ID, and server X-Request-ID. Transport, protocol, wait-timeout, terminal-operation, pagination, and integrity failures are distinct. Exceptions do not include keys, request bodies, queries, raw HTML, or document content.

Optional reranking (0.2.0)

After a project owner configures their paid Jev key in the dashboard, request reranking for an individual search. This interface requires ragwell>=0.2.0:

from ragwell import Ragwell
from ragwell.types import RerankRequest

with Ragwell() as client:
    result = client.project(project_id).search(
        query="What is the refund window?",
        k=5,
        rerank=RerankRequest(id="jev"),
    )
    for item in result.items:
        print(item.scores.rerank, item.content)

AsyncRagwell exposes the same argument on await project.search(...). Omitting rerank or passing None uses ordinary retrieval. A missing project key raises ConflictError with code project_reranker_misconfigured. Provider failures remain explicit, and the SDK never automatically retries a search. Jev charges the configured TypeSafe, OpenRouter or Vercel AI Gateway account. The client API key is your Ragwell key; never pass the provider key as a search parameter. Provider, API base URL and Jev model are project settings; the Ragwell SDK request stays the same for every route. response.rerank includes the provider, requested model and resolved_model (null when no scoring ran), so gateway model aliases remain traceable.

Relevance is 0–3; rerank_confidence is separate on 0–1. Original source parts, text/vector scores and hybrid score remain available; scores.final follows the reranking order. The first version accepts no tuning parameters.

Runnable async example

The Python 3.12 example project installs the published SDK from PyPI. Copy its env.example to .env, configure your project, then run its scripts to sync a local document folder, retrieve to terminal/JSON, and clean the project.

Development and qualification

uv run --locked ruff format --check .
uv run --locked ruff check .
uv run --locked mypy
uv run --locked python scripts/generate.py --check
uv run --locked pytest
uv run --locked python -m build --no-isolation

The vendored artifact is 2026-09-22.1, machine SHA-256 edd7aa3d26def4a1dcafceee4af2b9d4ae5a0d2affde52719ce03e904cc5a509. See contract provenance, generation qualification, and operation mapping. Ordinary tests require no service, credentials, provider calls, private repository, or model downloads.

The current contract adds optional Jev reranking. The installed 0.2.0 candidate passed the deployed beta lifecycle with the matching live contract, and its package runtime passed a bounded synthetic Jev request. Representative quality, latency and cost evaluation remains separate product-rollout work. The release process records the remaining hosted CI, TestPyPI and PyPI controls.

An optional beta lifecycle validation checks the installed wheel against a dedicated existing beta project with a scoped key. It starts no services and is separate from the full fixture matrix and ordinary offline CI.

Documentation: API reference and scopes, sync lifecycle example, async lifecycle example, compatibility and support, security policy, changelog, and release preparation.

License

MIT.