feat: Add baseten support

[monsieurBoutte]

Overview

Adds Baseten as a fourth LLM/embedding provider (provider: "baseten") across language-model workflows and generateEmbeddings. Baseten has no baked-in default model — callers must pass model or set BASETEN_MODEL / BASETEN_EMBEDDING_MODEL. Language models can use Baseten Model APIs or a dedicated OpenAI-compatible /sync/v1 deployment URL; embeddings require a dedicated /sync or /sync/v1 URL. Baseten is wired for runtime use but excluded from eval runs for now. Summarization with Baseten is transcript-only (no storyboard frames), even for video assets.

What was changed

• src/lib/providers.ts — Core integration: SupportedProvider / SupportedEmbeddingProvider extended; Baseten client creation via @ai-sdk/baseten; URL resolution for dedicated deployments; getDefaultLanguageModel / getDefaultEmbeddingModel helpers; eval model lists skip Baseten.
• src/env.ts, src/types.ts, src/lib/workflow-credentials.ts — New Baseten env vars and per-request credential fields (basetenApiKey, basetenModelUrl, basetenEmbeddingModelUrl, etc.).
• src/workflows/summarization.ts — Baseten forces transcript-only analysis; validates transcript is present for video assets; audio-only prompts generalized from "audio content" to "transcript content".
• package.json — Adds @ai-sdk/baseten.
• tests/unit/providers-baseten.test.ts — Unit tests for model resolution, eval exclusion, credential/URL wiring, and embedding defaults.
• scripts/post-evalite-results.ts — Explicit error if Baseten is requested for eval insights.
• Docs (README.md, docs/API.md, docs/CREDENTIALS.md, docs/EVALS.md, docs/WORKFLOWS.md) — Baseten listed as a supported provider; credential setup documented; eval exclusion noted.
• Examples — --provider baseten added to CLI examples; DEFAULT_MODELS made Partial where Baseten has no hardcoded default.

Suggested review order

1. src/lib/providers.ts — Provider resolution, URL normalization, and model factory logic; this is the heart of the change.
2. src/workflows/summarization.ts — Transcript-only path for Baseten and the validation/error when no transcript is available on video assets.
3. tests/unit/providers-baseten.test.ts — Confirms expected resolution, eval filtering, and credential/URL behavior.
4. src/env.ts, src/types.ts, src/lib/workflow-credentials.ts — Credential surface area and env var naming.
5. scripts/post-evalite-results.ts — Eval exclusion guard.
6. Docs and examples — Mostly mechanical baseten additions; spot-check docs/CREDENTIALS.md and README Baseten setup section for accuracy.

---

Note

Medium Risk
Touches core provider resolution and credential paths used by many workflows; Baseten summarization behavior change (transcript-only for video) and deployment URL requirements can surprise callers if misconfigured.

Overview
Adds Baseten as provider: "baseten" for language-model workflows and generateEmbeddings, wired through @ai-sdk/baseten with new env vars, runtime credential fields, and URL resolution for Model APIs vs dedicated /sync / /sync/v1 deployments. Baseten has no library default model—callers must set model or BASETEN_MODEL / BASETEN_EMBEDDING_MODEL; embeddings require a dedicated deployment URL.

Summarization with Baseten uses transcript-only analysis (no storyboard), including for video assets, and errors when no transcript is available. Audio-only summarization prompts are generalized from “audio content” to “transcript content.”

Baseten is excluded from eval model iteration (filtered from MUX_AI_EVAL_MODELS, not in default/all sets); post-eval insights reject Baseten explicitly. Docs, examples, and API types list Baseten alongside existing providers; unit tests cover model resolution, URL normalization, and eval filtering.

^{Reviewed by Cursor Bugbot for commit ac90125. Bugbot is set up for automated code reviews on this repo. Configure here.}

[snyk-io]

✅ Snyk checks have passed. No issues have been found so far.

Status	Scan Engine	Critical	High	Medium	Low	Total (0)
✅	Open Source Security	0	0	0	0	0 issues
✅	Licenses	0	0	0	0	0 issues

💻 Catch issues earlier using the plugins for VS Code, JetBrains IDEs, Visual Studio, and Eclipse.