ValidationError: Unprocessable Entity

[Craigson]

Issue

Subscribing to certain models throws an Unprocessable Entity error with no more information or context.

Model

fal-ai/kling-video/v1.6/pro/text-to-video

Dependencies

"typescript": "^5",
"@fal-ai/client": "^1.5.0",

Reproduction

This is the input passed to the model:

{
              "cfg_scale": 0.5,
              "duration": "5",
              "negative_prompt": "blur, distort, and low quality",
              "prompt": "A medieval knight pulling a magic sword from a stone",
              "aspect_ratio": "16:9"
}

Using the example same input via the Fal playground UI does not throw the error.

import { fal } from "@fal-ai/client";

export class FalModelCaller implements BaseModelCaller {
  private fal: typeof fal;

  constructor(apiKey: string) {
    this.fal = fal;
    this.fal.config({
      credentials: apiKey,
    });
  }

  // other methods ommited for brevity

  async callModel(params: {
    modelName: string;
    useCase: ModelUseCase;
    input: Record<string, any>;
    logs?: boolean;
    onUpdate?: (update: any) => void;
  }): Promise<AssetResult[]> {
    const { modelName, input, logs = true, onUpdate } = params;
    const pollInterval =
      params.useCase === MODEL_USE_CASES.IMAGE_GENERATION ? 500 : 2500;

    try {

      const result = await this.fal.subscribe(modelName, {
        input,
        logs,
        mode: "polling",
        pollInterval,
        onQueueUpdate: (update) => {
          switch (update.status) {
            case "IN_QUEUE":
              console.log("Queued");
              break;
            case "IN_PROGRESS":
              console.log("in progress");
              onUpdate?.(update);
              if (update.logs) {
                update.logs.map((log) => log.message).forEach(console.log);
              }
              break;
            case "COMPLETED":
              console.log("completed fal model execution");
              break;
          }
        },
      });

      // NOTE: <==== Error is thrown before reaching this point

      const responseAdapter = this.getResponseAdapter(params.useCase);
      const processedResult = responseAdapter.transform(result.data);
      return processedResult;
    } catch (error) {
      console.error("Error calling FAL model:", error);
      throw error;
    }
}

Error

This is the error thrown by the Fal client:

    {
      "message": [
        "Error calling FAL model:",
        "ValidationError: Unprocessable Entity"
      ],
      "level": "error",
      "timestamp": 1753117956817
    },

[Craigson]

@efiop @drochetti

I received this feedback via support email: The input should be body and in json format..

As per the documentation at migration and api the input param does not require a body property.... or perhaps I'm missing something.

From the documentation:

import { fal } from "@fal-ai/client";

const result = await fal.subscribe("fal-ai/kling-video/v1.6/pro/text-to-video", {
  input: {
    prompt: "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse."
  },
  logs: true,
  onQueueUpdate: (update) => {
    if (update.status === "IN_PROGRESS") {
      update.logs.map((log) => log.message).forEach(console.log);
    }
  },
});
console.log(result.data);
console.log(result.requestId);

Additionally, I get both the in progress and completed fal model execution logs, so the model execution appears to be receiving the input and processing it. My FalModelCaller is a generic class that is used to call all Fal model's, and ~90% execute without issue. If it was an issue with the input, I'd expect it to affect all model calls?

[Craigson]

It seems like this is the most likely culprit for the issue, from here

Warning: We are currently in a migration process and some APIs might not strictly follow the error structure documented below. We’re working to ensure all errors conform to this specification as soon as possible.

The docs state that an error object is returned, which I would then expect the following implementation to hold true:

    try {
      let result: any;

      try {
        result = await this.fal.subscribe(modelName, {
          input,
          logs,
          mode: "polling",
          pollInterval,
          onQueueUpdate: (update) => {
            switch (update.status) {
              case "IN_QUEUE":
                console.log("Queued");
                break;
              case "IN_PROGRESS":
                // console.log("in progress");
                // onUpdate?.(update);
                // if (update.logs) {
                //   update.logs.map((log) => log.message).forEach(console.log);
                // }
                break;
              case "COMPLETED":
                console.log("completed fal model execution");
                break;
            }
          },
        });
        if (
          result?.body &&
          result?.body?.detail &&
          Array.isArray(result.body.detail)
        ) {
          const errorDetail = result.body.detail[0];
          if (errorDetail && errorDetail.type === "content_policy_violation") {
            throw new Error(`Content policy violation: ${errorDetail.msg}`);
          }
          // Handle other error types if needed
          throw new Error(
            `FAL API error: ${errorDetail?.msg || "Unknown error"}`
          );
        }
      } catch (error) {
        console.error("calling the actual model failed:", error);
        

      }

      console.log({ falResult: result });

However, instead of returning the error object as the result, as described, an error is thrown and subsequently caught by calling the actual model failed:.

Reviewing the model execution via the Fal analytics UI, it shows the output as:

{
  "detail": [
    {
      "loc": [
        "body"
      ],
      "msg": "The content could not be processed because it contained material flagged by a content checker.",
      "type": "content_policy_violation",
      "url": "https://docs.fal.ai/errors#content_policy_violation",
      "input": {
        "prompt": "Donald trump eating a sandwich",
        "duration": "5",
        "aspect_ratio": "16:9",
        "negative_prompt": "blur, distort, and low quality",
        "cfg_scale": 0.5
      }
    }
  ]
}

Which is what I'd expect to receive as the result.

[Craigson]

Presumably this should return the error description as details instead of throwing:

fal-js/libs/client/src/response.ts

Line 71 in e10ba6d

throw new ErrorType({

[drochetti]

Hey @Craigson, apologies for the delay here.

Throwing the error is expected with the current design on the client. Returning error as the result, although a valid design, would break the current behavior for existing customers.

When it comes to validation errors, if you check the error for error instanceof ValidationError you should be able to get all the details, including which fields failed in a typesafe manner.

Have you tried that?

[Kelier]

i found a new error about https://fal.ai/models/fal-ai/wan/v2.2-a14b/image-to-video/api ,if you add end_image_url, you'll get the same situation