Chat and streaming

Every model you add is reachable at <your api>/ai/v1, in OpenAI’s formats. The SDKs wrap it as pawabase.ai.

Servers use a secret key and can call any model. A browser can use only the public key, and only models set to Signed-in users or Anyone. See Providers and models.

Ask a model

const text = await pawabase.ai.chat("chat", "Summarise this ticket: …");

// a whole conversation, with options
const reply = await pawabase.ai.chat(
  "chat",
  [
    { role: "system", content: "Answer in one sentence." },
    { role: "user", content: "What is a vector database?" },
  ],
  { temperature: 0.2, maxTokens: 120 },
);

// everything the model returned: tool calls, finish reason, token usage
const completion = await pawabase.ai.complete("chat", "Hello");
completion.usage?.total_tokens;

text = pb.ai.chat("chat", "Summarise this ticket: …")

reply = pb.ai.chat(
    "chat",
    [{"role": "system", "content": "Answer in one sentence."}, {"role": "user", "content": "What is a vector database?"}],
    temperature=0.2, max_tokens=120,
)

completion = pb.ai.complete("chat", "Hello")
completion["usage"]["total_tokens"]

Stream the answer

for await (const piece of pawabase.ai.stream("chat", "Tell me a short story")) {
  process.stdout.write(piece);
}

for piece in pb.ai.stream("chat", "Tell me a short story"):
    print(piece, end="", flush=True)

# async client
async for piece in pb.ai.stream("chat", "Tell me a short story"):
    ...

A failure before the first word is raised as an error with the provider’s reason. A failure partway through ends the stream with an error.

Tools and JSON

Pass tools in OpenAI’s format to let a model ask you to call a function, and responseFormat: { type: "json_object" } (response_format= in Python) to ask for JSON. Anthropic’s native format is translated for you, so the same request works for every provider.

const completion = await pawabase.ai.complete("chat", "What's the weather in Lagos?", {
  tools: [{ type: "function", function: { name: "get_weather", description: "Current weather",
    parameters: { type: "object", properties: { city: { type: "string" } }, required: ["city"] } } }],
});
const call = completion.choices[0].message.tool_calls?.[0];   // run it yourself, then send the result back as a "tool" message

Pawabase does not run the tool loop for you: you call the function and send its result back in the next request.

Any OpenAI client

import OpenAI from "openai";

const ai = new OpenAI({
  baseURL: "https://api.example.com/ai/v1",
  apiKey: "unused",
  defaultHeaders: { apikey: process.env.PAWABASE_SECRET_KEY },
});
const reply = await ai.chat.completions.create({ model: "chat", messages: [{ role: "user", content: "Hello" }] });

Errors

Gateway errors use OpenAI’s shape, { "error": { "message", "type", "code" } }, and the SDKs raise them with the message and code kept.

Status	`code`	Meaning
401	`unauthenticated`	The model needs a signed-in user
403	`forbidden`	The model’s rule refused this caller
404	`model_not_found`	No such model, or it is switched off
400	`wrong_kind`	A chat model was used for embeddings, or the reverse
422	`bad_request`	Missing `model`, empty `messages`, or a bad `input`
502 / 504	`provider_error`	The provider failed or did not answer. A fallback model, if set, is tried first.

Getting started

Models

Agents

Knowledge bases

Connections and tools

Help

Chat and streaming

Ask a model

Stream the answer

Tools and JSON

Any OpenAI client

Errors

​Ask a model

​Stream the answer

​Tools and JSON

​Any OpenAI client

​Errors

Ask a model

Stream the answer

Tools and JSON

Any OpenAI client

Errors