Mistral

Mistral file-backed text and embedding batch jobs in Batchwork with normalized results and the shared async client.

Mistral uses file-backed Batch Jobs for chat and embeddings. Its JSONL and create payload differ from OpenAI: each line contains only custom_id and body, while the endpoint and model are supplied when the job is created.

At a glance

Property	Batchwork behavior
Output modalities	Text and embeddings
Text inputs	Messages, tools, images, and PDFs
Endpoints	`/v1/chat/completions`, `/v1/embeddings`
Submission	JSONL upload, then `/batch/jobs`
Results	Output and error JSONL files
Batch metadata	Forwarded
Webhooks	Managed polling only
Credential	`MISTRAL_API_KEY`
Default base URL	`https://api.mistral.ai/v1`

Text example

from batchwork import BatchRequest, Batchwork

async with Batchwork() as client:
    job = await client.batch(
        model="mistral/mistral-large-latest",
        requests=[BatchRequest(custom_id="summary", prompt="Summarize this report.")],
    )

Embeddings

from batchwork import BatchEmbeddingRequest

job = await client.batch_embeddings(
    model="mistral/mistral-embed",
    requests=[BatchEmbeddingRequest(custom_id="doc-1", value="Document text")],
)

Mistral does not support canonical dimensions. Public Python and CLI submissions reject unknown Mistral embedding provider options before upload.

Lifecycle

Each JSONL line contains {custom_id, body}; model and stream are removed from the body.
Batchwork uploads the JSONL file.
/batch/jobs receives the file ID, endpoint, model, and optional batch metadata.
poll() retrieves /batch/jobs/{id}.
results() streams the output file and then the error file when present.
cancel() calls /batch/jobs/{id}/cancel.

Mistral-specific uppercase states, including timeout and cancellation-requested states, map to normalized BatchStatus values.

Request translation

Input	Mistral behavior
`seed`	Renamed to `random_seed`
Frequency/presence penalties	Ignored
Assistant reasoning	Folded into assistant text
Final assistant message	Receives `prefix=True`
Named tool choice	Filters tools to the named function and sends `tool_choice="any"`
Required tool choice	Becomes `"any"`

Mistral provider options include safe prompt, reasoning effort, document image/page limits, and parallel tool calls.

CLI capability contract

Exact text provider-option keys are documentImageLimit, documentPageLimit, parallelToolCalls, reasoningEffort, and safePrompt. Only chat-completions is a valid explicit --endpoint. Unknown keys fail locally; document limits are positive integers, parallelToolCalls and safePrompt are booleans, and reasoningEffort is a string. Canonical frequency/presence penalties and top_k are unsupported.

CLI embedding provider options: none. Canonical embedding dimensions is also unsupported and fails before upload. Batch metadata is supported and forwarded. Package hard bounds are 50,000 requests, 20 MiB per serialized request, and 200 MiB aggregate, subject to lower Mistral/model limits.