Groq

Provider groq. Adapter over the groq SDK, which speaks the same chat.completions dialect as OpenAI — that's why it inherits from _OpenAICompatible.

pip install "jangada-ai[groq]"

• provider=: "groq"
• Environment variable: GROQ_API_KEY
• Difference from OpenAI: supports_parse_helper = False (no native .parse()).

from jangada_ai import LLM
llm = LLM("groq", "llama-3.3-70b-versatile")

What it does

• Text and streaming — focused on very low latency.
• Structured output (parse): since there is no .parse helper, it uses response_format={"type":"json_schema",...} and validates with model_validate_json.
• Vision (images=): supported only on vision-capable models (e.g., the Llama vision family). Text-only models reject images.
• Documents (files=): local text extraction.
• Audio transcription (transcribe): dedicated endpoint (compatible with OpenAI's). Models: whisper-large-v3, whisper-large-v3-turbo (fast).

Structure and quirks

• Parameters: temperature, max_tokens, top_p, stop, seed. No top_k.
• Same base as OpenAI: message translation, structured, and audio reuse _OpenAICompatible; only sdk_module/sync_class/ async_class/supports_parse_helper change.
• Response (Completion): same as OpenAI (text, usage, raw).

When to choose Groq

Low inference cost and speed — great for batch transcription (whisper-large-v3-turbo) and low-latency responses. Combine it as a fallback or primary with OpenAI (same dialect). See Audio transcription and Retry and fallback.

What changed in 1.9.0

• Native tools: on groq/compound* models, web_search(), web_fetch() and code_execution() enable the built-in tools (compound_custom); on openai/gpt-oss-*, web_search() becomes browser_search and code_execution() becomes code_interpreter. Compound doesn't accept user function tools in the same call. See Native tools.
• Structured output: fields with defaults are now listed in required in strict mode (Groq was rejecting the schema), and the fallback to JSON Object mode triggers on more error messages.