Skip to content

Supported parameters

Flockadoo accepts an OpenAI-shaped request and forwards the parameters both vLLM and llama.cpp can run. A key that is not on the list for that endpoint returns 400. A parameter that is on the list but not enabled on the provider that can serve the model returns 400 capability_unavailable once every provider has been checked, or 503 while checks are still running.

Chat and completions

You can send model, messages or prompt, stream, max_tokens, max_completion_tokens, temperature, top_p, top_k, min_p, stop, seed, frequency_penalty, presence_penalty, tools, tool_choice, response_format, logprobs, top_logprobs, and stream_options. Chat requests can also send reasoning_effort (see Reasoning models).

n and best_of may be 1. Any larger value returns 400. Flockadoo bills one completion.

flockadoo, user, metadata, store, and service_tier are accepted and not forwarded to the engine.

A parameter sent as null is treated as if you left it out. seed must be a whole number.

Messages

Message content is text only. content can be a string or a list of parts whose type is text. An image, audio, or file part returns 400, even if the model could handle it.

Tools

tool_choice needs a non-empty tools list, unless it is "none". Sending "auto", "required", or a named function without tools returns 400.

tool_choice is "none", "auto", "required", or {"type": "function", "function": {"name": "..."}}. A named function must be in tools.

If you force a tool call ("required" or a named function), the tool's parameters schema must follow the same rules as a JSON schema below.

JSON schema output

response_format.type is text, json_object, or json_schema. A json_schema may use these keywords: type, properties, required, items, and enum. Any other keyword returns 400.

title, description, default, examples, and additionalProperties: false are also accepted, so schemas generated by libraries such as Pydantic work as-is. They are removed before the request reaches the engine and do not change the output. additionalProperties: true returns 400.

Reasoning models

Some models think before they answer: Qwen3, DeepSeek-R1, gpt-oss, and Gemma 4. Their thinking arrives apart from the answer, in its own field. Flockadoo passes the provider's engine output through unchanged, so the field's name depends on the engine: reasoning from vLLM, reasoning_content from llama.cpp. Neither is part of OpenAI's schema; treat both as optional provider extensions.

If you only want the answer, read content as usual; the thinking is not in it. Flockadoo serves a reasoning model only from providers whose engine was checked to send the thinking apart from the answer. While that check is still running for every provider of the model, the request returns 503; if every provider failed it, 400 capability_unavailable.

To show the thinking, read whichever field is present. In a stream:

const thinking = delta.reasoning ?? delta.reasoning_content;

In a response that is not streamed, read the same fields from message:

const thinking = message.reasoning ?? message.reasoning_content;

With the OpenAI Python SDK the fields are extra attributes:

thinking = getattr(delta, "reasoning", None) or getattr(delta, "reasoning_content", None)

Thinking is generated output. It starts a stream, keeps it inside the time limits below, counts toward the output limit, and is billed as output.

reasoning_effort is none, minimal, low, medium, high, xhigh, or max. none turns thinking off. Any other value turns it on for a model that only thinks when asked (Gemma 4). How much the other levels change the answer depends on the model. A level other than none on a model that does not reason returns 400 capability_unavailable. Any other value, or reasoning_effort on /v1/completions, returns 400.

Embeddings

Send model and input. encoding_format may be float or base64. user is accepted and not forwarded. dimensions and generation parameters return 400.

A batch may contain at most 256 inputs. An empty string is rejected. Each input must fit the model's context length. The batch must stay within 300,000 tokens.

Time limits

A non-streaming request has 5 minutes for the provider to finish. If that deadline hits while you are still connected, the response is 504 with code inference_deadline_exceeded and message provider did not finish within the non-streaming time limit.

A streaming request has 2 minutes to produce the first generated output (a reasoning model's thinking counts), then 1 minute between later generated chunks, and 30 minutes in total.

If nothing was delivered, you are not billed. After the first generated output is delivered, you pay for the input once and for the output that was delivered.

Output length

If you omit both token limits, the platform cap is 16,384 generated tokens. If you set either limit, Flockadoo uses the smaller of the values you set, and never more than 16,384. Only max_tokens is forwarded.