Cloudflare provider analyzeImage() returns empty string for Workers AI vision models

[stackbilt-admin]

Summary

LLMProviders.analyzeImage() with the cloudflare provider returns { content: "", message: "" } when called from a Cloudflare Worker binding. The vision step completes without throwing, but produces no usable text, silently breaking downstream consumers.

Reproduction

Context: foodfiles POST /v2/recipes/analyze (apps/api/src/routes/recipes.ts)

const llm = new LLMProviders({
  cloudflare: { ai: c.env.AI },
  defaultProvider: "cloudflare",
  costOptimization: true,
  enableCircuitBreaker: true,
});
const visionResult = await llm.analyzeImage({
  image: { data: base64, mimeType: imageFile.type },
  prompt: "Describe this food image in detail...",
  maxTokens: 512,
});
// visionResult.content === ""
// visionResult.message === ""

Model selected by getDefaultVisionModel() for cloudflare: @cf/meta/llama-3.2-11b-vision-instruct

Root cause hypothesis

attachImagesToLastUserMessage() (cloudflare.ts) formats the image as a { type: "image_url", image_url: { url: "data:image/jpeg;base64,..." } } content part in the messages array.

Workers AI's llama-3.2-11b-vision-instruct appears to return { choices: [{ message: { content: null } }] } for this format when called via the Workers binding (not the REST API). The extractText() null-content path returns "".

The Workers AI binding expects the raw format for vision:

ai.run('@cf/meta/llama-3.2-11b-vision-instruct', {
  image: [...], // number[] / Uint8Array
  prompt: "...",
  max_tokens: 512,
})
// → { response: "description text" }

The REST API and the binding have different input shapes for this model. The provider only implements the messages/image_url path.

Evidence

• Direct external call to tarotscript-worker.blue-pine-edf6.workers.dev/v2/recipes/analyze with a hand-crafted image_analysis string succeeds — confirming downstream is fine.
• Tarotscript returned 400 image_analysis is required when the llm-providers vision step was allowed to pass through its empty string, confirming visionResult.content and visionResult.message are both "".
• extractText() handles chatContent === null by returning "" (cloudflare.ts ~line 562) — no error thrown, silent failure.

Suggested fix

For Workers AI vision models, use the raw binding format:

ai.run(model, {
  image: Array.from(imageBytes), // number[]
  prompt: request.messages[lastUserIdx].textContent,
  max_tokens: request.maxTokens,
})

and map the { response: string } return back through formatResponse. The messages/image_url path can remain for REST API consumers.

Alternatively, detect at runtime whether the AI binding is present and branch the format accordingly.

Version

@stackbilt/llm-providers ^1.5.0 (foodfiles dependency)