@dbx-tools/model-protocol

Construct Vercel AI SDK ProviderV4 models for Databricks endpoints and translate gateway traffic between supported model protocols.

Invoke a Databricks model

Construct the AI SDK model after resolving the endpoint protocol and fresh authentication headers for the current request:

import { createDatabricksLanguageModel } from "@dbx-tools/model-protocol/provider";
import { generateText } from "ai";

const model = createDatabricksLanguageModel({
  host: process.env.DATABRICKS_HOST!,
  modelId: "databricks-gpt-5-6-sol",
  protocol: "responses",
  headers: { authorization: `Bearer ${process.env.DATABRICKS_TOKEN}` },
});

const result = await generateText({ model, prompt: "Summarize the incident." });
console.log(result.text);

Use this package when a Node service needs to:

• construct a request-scoped AI SDK model for Databricks Chat Completions, Responses, or Anthropic Messages.
• translate OpenAI or Anthropic gateway messages to and from AI SDK messages.
• normalize Databricks Chat Completions request and response wire shapes.

Model discovery and capability policy remain in @dbx-tools/model. Browser-safe model and gateway contracts remain in @dbx-tools/shared-model and @dbx-tools/shared-model-gateway.

Use chat, responses, or anthropic according to the model-owned protocol selection. The chat provider applies Databricks serving-wire repairs through a provider-local fetch. It does not patch globalThis.fetch.

Translate a gateway request

Decode an OpenAI or Anthropic request into AI SDK messages and encode the completed generation back into the caller's protocol:

import { decodeGatewayRequest } from "@dbx-tools/model-protocol/gateway-decode";
import { encodeGatewayResponse } from "@dbx-tools/model-protocol/gateway-encode";
import { createDatabricksLanguageModel } from "@dbx-tools/model-protocol/provider";
import { generateText } from "ai";

const protocol = "openai-chat" as const;
const decoded = decodeGatewayRequest(protocol, {
  messages: [{ role: "user", content: "Hello" }],
});
const model = createDatabricksLanguageModel({
  host: process.env.DATABRICKS_HOST!,
  modelId: "databricks-gpt-5-6-sol",
  protocol: "responses",
  headers: { authorization: `Bearer ${process.env.DATABRICKS_TOKEN}` },
});
const generated = await generateText({ model, ...decoded });

const response = encodeGatewayResponse(protocol, "databricks-gpt-5-6-sol", {
  text: generated.text,
  toolCalls: generated.toolCalls,
  usage: generated.usage,
  finishReason: generated.finishReason,
});

Use encodeGatewayStream() instead when forwarding streamText().fullStream. The decoder and encoders share tool-call, reasoning, usage, and finish-reason handling between Mastra and the model gateway.

Protocol coverage and limits

The provider side supports Databricks Chat Completions, Responses, and Anthropic Messages. The gateway decoder and encoders support OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages for buffered and streaming text and function-tool traffic. The model gateway sends embeddings directly and doesn't translate them through this package.

Fallback translation preserves text, function tools, tool results, reasoning events, token usage, and finish reasons. It doesn't implement stateful Responses features such as previous_response_id, store: true, or background: true.

The Chat provider's request-scoped fetch removes unsupported fields, applies GPT 5.6 tool-call reasoning compatibility, strips provider reasoning blocks that can't be replayed, repairs assistant prefill after tool results, and flattens array-valued response content while preserving returned thinking as AI SDK reasoning events. It doesn't modify globalThis.fetch.