Videodb Query

result = videodb_query_skill("get_contexts")
# Returns: selected_videos, selected_pdfs, num_videos, num_pdfs

2. get_summary

Get overview statistics for selected videos.

result = videodb_query_skill("get_summary")
# Returns: video names, durations, object counts per video

3. semantic_search

Search video segments using natural language descriptions.

result = videodb_query_skill("semantic_search", query="convoy moving through forest")
# Returns: matching segments with similarity scores

4. object_search

Find segments containing specific object types.

Valid object types: tank, truck, soldier

result = videodb_query_skill("object_search", object_type="tank")
# Returns: timeline_text (pre-formatted), summary_text, max_objects_at_any_point

Important: Use timeline_text and summary_text directly in your response. Never sum object counts across segments - the same object may appear in multiple consecutive segments.

5. event_search

Search by event keywords in AI-generated descriptions.

result = videodb_query_skill("event_search", query="moving")
# Returns: segments with matching descriptions

6. segment_details

Get detailed information about a specific segment.

result = videodb_query_skill("segment_details", segment_id=0, video_ids=["m-abc123"])
# Returns: event_description, object_counts, individual object positions

Temporal Continuity Awareness

Critical Understanding: The same object appearing in consecutive segments is NOT multiple objects. When reporting object counts:

• WRONG: "Found 45 tanks" (summing all detections)
• CORRECT: "Maximum 15 tanks observed at any point, appearing across 3 segments"

The object_search action returns pre-formatted timeline_text that handles this correctly. Use it directly.

Multi-Video Analysis

When multiple videos are selected:

1. Assess relationships: Are videos sequential (same mission) or independent?
2. Attribution: Always specify which video findings come from
3. Counting methodology: Related videos may show same objects

Example Multi-Video Workflow

# Get contexts
contexts = videodb_query_skill("get_contexts")

# Query each video
for video_id in contexts["result"]["selected_videos"]:
    result = videodb_query_skill("object_search", object_type="tank", video_ids=[video_id])
    # Process results per video

# Provide context-aware response

Result Structure

All actions return:

• status: "success", "no_results", or "error"
• action: The action performed
• result: Action-specific data
• message: Human-readable summary

Best Practices

1. Always call get_contexts first
2. Check status before using results
3. Use timeline_text directly for object counts
4. Explain your counting methodology
5. Attribute findings to specific videos
6. Handle "no_results" gracefully (it's not an error)