Mistral

Provider mistral. Adapter over the official mistralai SDK (not the OpenAI-compatible path). It gives you Mistral's models through the same LLM(...), with text, vision, streaming, structured output, function calling, embeddings, OCR/Document AI and audio transcription (Voxtral).

pip install "jangada-ai[mistral]"

• provider=: "mistral"
• Authentication: the MISTRAL_API_KEY environment variable (or api_key= in the constructor).
• Model: a Mistral id, e.g. mistral-large-latest, mistral-medium-latest, mistral-small-latest, ministral-8b-latest, codestral-latest, mistral-embed, mistral-ocr-latest, voxtral-mini-latest.

from jangada_ai import LLM

llm = LLM("mistral", "mistral-large-latest")
resp = llm.complete("Summarize {{ topic }} in one sentence.", topic="rafts")
print(resp.text, resp.cost)

Configuration via environment

In .env (or process variables):

MISTRAL_API_KEY=...

What it does

• Text and streaming via chat.complete/chat.stream (and async variants).
• Structured output (parse): native chat.parse(response_format=Model) helper — takes the Pydantic model directly and returns message.parsed.
• Tools (tools=/tool_choice=): OpenAI format; tool_choice accepts auto/none/any/required or a function name (forces the call).
• Vision (images=): multimodal models (Pixtral/medium/large).
• Embeddings (embed/aembed) with mistral-embed.
• OCR / Document AI (ocr/aocr) with mistral-ocr-latest: PDF/image → per-page markdown, bounding boxes and extracted images.
• Transcription (transcribe/atranscribe) with Voxtral.

# OCR: accepts a URL, path, bytes, an ImagePart or an API `document` dict.
ocr = LLM("mistral", "mistral-ocr-latest")
doc = ocr.ocr("https://arxiv.org/pdf/2310.06825.pdf", include_images=True)
print(len(doc.pages), doc.pages[0].markdown)

# Transcription (Voxtral).
stt = LLM("mistral", "voxtral-mini-latest")
print(stt.transcribe("audio.mp3", language="en").text)

Quirks

• Mistral uses random_seed (jangada maps the canonical seed) and has no top_k (dropped). Streaming deltas come in event.data.choices[0].delta.content.
• Server-side MCP is not supported here — use the client-side MCPClient or the Anthropic/OpenAI/Groq providers for remote MCP by URL.

See Providers and Capabilities matrix.

What changed in 1.9.0

• Native tools via the Conversations API: web_search() (or web_search(premium=True)), code_execution(), image_generation() and file_search(stores=[...]) (document library) route the call through beta.conversations. User function tools work alongside; with native tools, stream and parse are not supported. With params={"store": True}, the next turn continues the conversation server-side. See Native tools.
• Batched embeddings: large lists are split into several requests.
• An empty response becomes ServerError; tool arguments with invalid JSON are kept in metadata["raw_arguments"].