# flexiclass

Text and image classification, evidence and verification. Ask one question or several about the same documents. Use the origin of this URL as the API base URL.

## Documentation

- [Overview](/human): product explanation and checked examples.
- [Integration guide](/clanker.md): complete Markdown instructions, Python CSV workflow, batching and failures.
- [OpenAPI schema](/openapi.json): authoritative request/response schema and configured limits.
- [Python client](/clanker/client.py): standard-library helpers classify_rows, classify_groups and ask_rows; classify_images and ask_images for images; rank_candidates and retrieve_classify for retrieval; create_embeddings for power users; save as classify_rows.py.
- [Browser guide](/clanker): readable HTML when Accept is text/html.
- [Interactive reference](/docs): explore and try API calls.

## Access

The operator supplies FLEXICLASS_API_KEY. In protected mode, send Authorization: Bearer <FLEXICLASS_API_KEY> when fetching the guide, schema or client. Use Accept: text/plain for the Python client. Never put credentials in URLs or source code. A 401 response means credentials are missing or invalid, not that the file is absent.

## Start

Read the integration guide and schema before writing code. POST /v1/flexiclass accepts one text or a texts list, and either one question (labels, with an optional task and descriptions) or several (questions, each with its own task and labels). Every question is asked of every document and answered independently; asking several together is faster on longer documents, where it can move individual answers on ambiguous questions. Group rows that share questions, preserve IDs locally and inspect every result status. Every answer carries reliability, the probability that it is right; use confidence to decide what can proceed and what needs review. Results carry candidate scores unless include_scores is false. Write labels as meaningful answers: for two options, use the condition and its negation, not yes/no. effort accepts low for speed, medium for balance, or high for more care. GET /v1/flexiclass describes the service and its limits.

## Images

POST /v1/flexiclass/image takes one image or an images list (each {data: base64 JPEG, PNG or WebP, optional media_type, optional text}) with the same labels or questions and returns the same response. There is no detail setting: every image is read at one fixed resolution, and the service scales larger images down itself. The downloadable client defaults to 4 images per request with up to 4 requests in flight. The recent local screen also supports trying single-image requests with 16 workers on the reference deployment. Image answers carry reliability too; read it the same way. Images support low effort only; GET /v1/flexiclass/image reports enabled, available and the image limits. See the Images section of the integration guide.