Skip to content

Add Databricks provider (AI Gateway / Foundation Model APIs)#1433

Draft
prasadkona wants to merge 2 commits intoanomalyco:devfrom
prasadkona:add-databricks-ai-gateway-provider
Draft

Add Databricks provider (AI Gateway / Foundation Model APIs)#1433
prasadkona wants to merge 2 commits intoanomalyco:devfrom
prasadkona:add-databricks-ai-gateway-provider

Conversation

@prasadkona
Copy link
Copy Markdown

@prasadkona prasadkona commented Apr 13, 2026

Summary

Adds a Databricks catalog under providers/databricks/ for Foundation Model APIs exposed through Databricks AI Gateway: OpenAI-compatible chat/embeddings on mlflow/v1, plus per-model overrides for Anthropic Messages, Gemini generateContent, and OpenAI Responses—all on the same ai_gateway_url host returned by workspace discovery.

What’s included

  • providers/databricks/provider.toml, logo.svg, per-model models/*.toml (capabilities, limits, costs where applicable, [provider] overrides for non-MLflow surfaces).
  • providers/databricks/README.md — Scope, discovery (GET /api/ai-gateway/v2/endpoints), HTTP layout, auth, eligibility rules, maintainer scripts.
  • Root README.md — One-line pointer to the Databricks provider doc.
  • Scripts (repo root package.json):
    • bun run databricks:list-gateway — list/filter gateway routes (discovery helper).
    • bun run databricks:test-inference — E2E against live gateway (packages/core/script/test-databricks.ts).
  • packages/core/script/databricks-ai-gateway-shared.ts — Shared discovery + URL helpers for list/test scripts.
  • packages/core/package.json@databricks/sdk-experimental for discovery/auth in scripts.

Validation / how reviewers can check

bun install
bun validate

Optional (requires Databricks auth + workspace with AI Gateway routes):

bun run databricks:list-gateway -- --profile <profile>
bun run databricks:test-inference -- --profile <profile>

Testing

  • bun validate passes locally
  • databricks:test-inference exercised against a workspace profile (36/36 gateway routes) where applicable

Notes

  • Catalog entries are metadata only; no secrets in-repo. Runtime clients must perform discovery and attach a Bearer token.

@prasadkona prasadkona force-pushed the add-databricks-ai-gateway-provider branch from c791a73 to 57fd9ae Compare April 13, 2026 14:20
Catalog providers/databricks for Foundation Model APIs on AI Gateway: default mlflow/v1 via @databricks/ai-sdk-provider; per-model overrides for Anthropic Messages, Gemini generateContent, and Openai Responses (codex). Schema allows @databricks/ai-sdk-provider where api is required. Adds discovery and HTTP E2E scripts, README, and root README pointer.
@prasadkona prasadkona force-pushed the add-databricks-ai-gateway-provider branch from 57fd9ae to 1f7590b Compare April 13, 2026 23:43
Copy link
Copy Markdown

@anthonyivn2 anthonyivn2 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fields to review

- Correct attachment flags: gpt-oss-120b, gpt-oss-20b, qwen3-next-80b set to false
  (gateway returns "Image input is not supported for this endpoint")
- Add pdf to modalities.input for all Gemini models and Codex models
  (confirmed via inlineData / Responses API input_file probes)
- Set temperature=false for all gpt-5* models (policy rule):
  base variants return HTTP 400 "Only the default (1) value is supported";
  all gpt-5* unified to false for catalog consistency
- Fix gpt-5-nano tool_call: true (false negative - now uses tool_choice:"required")
- Add reasoning=true for gpt-5-nano (reasoning_tokens confirmed in usage)
- Add reasoning=true for gemma-3-12b (<think> blocks confirmed via system prompt)
- Add probe script: packages/core/script/probe-databricks-capabilities.ts
  Tests tool-use, reasoning, image, PDF, and temperature across all 4 API surfaces
  (mlflow/v1 chat, Anthropic, Gemini, OpenAI Responses). Read-only diagnostic tool.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants