ATH-MaaS/Pixelle-Video · error · KeyError

Invalid response format: missing 'video_prompts'

Error message

Invalid response format: missing 'video_prompts'

What it means

generate_video_prompts expects the LLM to answer each batch with JSON containing a 'video_prompts' key. If the parsed JSON lacks that key, a KeyError('Invalid response format: missing 'video_prompts'') is raised. Unlike the image variant it is raised immediately per attempt (no graceful count-retry messaging), so the exception surfaces whenever the model ignores the required schema.

Source

Thrown at pixelle_video/utils/content_generators.py:432

                prompt = build_video_prompt_prompt(
                    narrations=batch_narrations,
                    min_words=min_words,
                    max_words=max_words
                )
                
                response = await llm_service(
                    prompt=prompt,
                    temperature=0.7,
                    max_tokens=8192
                )
                
                logger.debug(f"Batch {batch_idx} attempt {attempt}: LLM response length: {len(response)} chars")
                
                # Parse JSON
                result = _parse_json(response)
                
                if "video_prompts" not in result:
                    raise KeyError("Invalid response format: missing 'video_prompts'")
                
                batch_prompts = result["video_prompts"]
                
                # Validate batch result
                if len(batch_prompts) != len(batch_narrations):
                    raise ValueError(
                        f"Prompt count mismatch: expected {len(batch_narrations)}, got {len(batch_prompts)}"
                    )
                
                # Success - add to all_prompts
                all_prompts.extend(batch_prompts)
                logger.info(f"✓ Batch {batch_idx} completed: {len(batch_prompts)} video prompts")
                
                # Report progress
                if progress_callback:
                    completed = len(all_prompts)
                    total = len(narrations)
                    progress_callback(completed, total, f"Batch {batch_idx}/{len(batches)} completed")

View on GitHub (pinned to 848b054e4f)

Solutions

1. Retry the batch generation — model nondeterminism often produces valid JSON on a later attempt.
2. Log/inspect the raw response and tighten the prompt to require the exact key "video_prompts".
3. Enable provider JSON mode or a response schema forcing the key.
4. Accept a top-level array as fallback before raising.
5. Use a stronger model for video prompt generation.

Example fix

// before
result = _parse_json(response)
if "video_prompts" not in result:
    raise KeyError("Invalid response format: missing 'video_prompts'")
// after
result = _parse_json(response)
if "video_prompts" not in result:
    if isinstance(result, list):
        result = {"video_prompts": result}
    else:
        raise KeyError("Invalid response format: missing 'video_prompts'")

def looks_like_video_prompt_response(payload) -> bool:
    return isinstance(payload, dict) and isinstance(payload.get("video_prompts"), list) and len(payload["video_prompts"]) > 0

def has_video_prompts(obj: object) -> bool:
    return isinstance(obj, dict) and isinstance(obj.get("video_prompts"), list) and all(isinstance(p, str) for p in obj["video_prompts"])

try:
    prompts = generator.generate_video_prompts(narrations)
except KeyError as e:
    if "video_prompts" in str(e):
        logger.warning("LLM returned unexpected shape; retrying")
        prompts = generator.generate_video_prompts(narrations)
    else:
        raise

When it happens

Trigger: Calling generate_video_prompts where _parse_json(response) yields a dict without 'video_prompts' — model returns prompts under another key, returns prose, or returns a top-level array.

Common situations: Model producing motion-caption-style text instead of JSON; schema instructions lost when customizing the prompt; provider switching to a model that resists JSON formatting; responses truncated so JSON parsed to an empty/partial object.

Related errors

• Invalid response format: missing 'narrations' key
• Invalid response format: missing 'image_prompts'
• Failed to parse LLM response as {response_type.__name__}: {c
• str(e)
• str(e)

AI-assisted analysis of ATH-MaaS/Pixelle-Video@848b054e4f (2026-08-30). Data as JSON: /api/errors/52c650353eb5588e. Report an issue: GitHub.