Initial draft of LangChain TypeScript Integration for Databricks Model Serving

[eLysgaard]

This PR introduces @databricks/langchain-ts, a TypeScript package that integrates Databricks Model Serving with the LangChain JS SDK. This is the TypeScript equivalent of the existing Python databricks-langchain integration.

Features

• ChatDatabricks - Core chat model implementation extending LangChain's BaseChatModel
• Streaming - Full SSE streaming support via _streamResponseChunks()
• Tool/Function Calling - bindTools() with support for OpenAI format tools, LangChain StructuredTools, and plain object definitions
• Unified Authentication - Uses @databricks/sdk-experimental for auth (PAT, OAuth M2M, Azure CLI/MSI, GCP credentials)
• Dual Module Format - ESM + CommonJS builds for broad compatibility

Usage

  import { ChatDatabricks } from "@databricks/langchain-ts";

  const model = new ChatDatabricks({
    endpoint: "databricks-meta-llama-3-3-70b-instruct",
  });

  // Basic invocation
  const response = await model.invoke("Hello!");

  // Streaming
  for await (const chunk of await model.stream("Tell me a story")) {
    process.stdout.write(chunk.content);
  }

  // Tool calling
  const modelWithTools = model.bindTools([...]);

  Package Structure

  integrations/langchain-js/
  ├── src/
  │   ├── index.ts              # Main exports
  │   ├── chat_models.ts        # ChatDatabricks implementation
  │   ├── types.ts              # OpenAI-compatible types
  │   ├── utils/
  │   │   ├── http.ts           # HTTP client with SSE streaming
  │   │   ├── messages.ts       # Message conversion utilities
  │   │   └── tools.ts          # Tool conversion utilities
  │   └── tests/
  │       ├── chat_models.test.ts      # Unit tests
  │       └── chat_models.int.test.ts  # Integration tests
  ├── examples/
  │   └── basic.ts              # Usage example
  ├── package.json
  └── README.md

Dependencies

Package	Purpose
@langchain/core (peer)	Base classes, message types
@databricks/sdk-experimental	Authentication
zod	Schema validation
zod-to-json-schema	Tool schema conversion

Test Plan

• Unit tests pass (npm test) - 16 tests covering message/tool conversion
• TypeScript type checking passes (npm run typecheck)
• ESLint passes (npm run lint)
• Build produces ESM + CJS outputs (npm run build)
• Example runs successfully against Databricks workspace (npm run example)
• Integration tests against live endpoint (npm run test:integration)

Future Work (not in this PR)

• DatabricksEmbeddings - Embeddings API support
• DatabricksVectorSearch - Vector search integration
• withStructuredOutput() - JSON mode / JSON schema support
• Responses API support for agent endpoints