entity-search

Entity Search - Find Specific People in Videos

Find specific people or objects in your indexed videos using reference images. Entity search lets you register a person with reference photos, then search for them across all your videos.

When to Use This Skill

Use this skill when the user:

• Wants to find a specific person in their videos
• Asks to set up face recognition or person tracking
• Has reference images of someone they want to locate in video content
• Wants to search for a known individual (e.g., "find all clips of John")

Prerequisites

• Videos must be indexed with Marengo 3.0 (entity search is a Marengo 3.0 feature)
• Reference images of the person/object (publicly accessible URLs or local file paths)
• Free plan: 1 collection, up to 15 entities. Developer plan: multiple collections.

Instructions

Step 1: Check for Existing Entity Collections

First check if the user already has entity collections set up:

Tool: mcp__twelvelabs-mcp__list-entity-collections

If collections exist, show them and ask if the user wants to use an existing one or create a new one.

Step 2: Create an Entity Collection (if needed)

An entity collection groups related entities. Create one per logical group (e.g., "Team A", "Film Cast").

Tool: mcp__twelvelabs-mcp__create-entity-collection
Parameters:
  name: "<collection name>"

Step 3: Upload Reference Images as Assets

Upload one or more reference images for the person. Multiple images from different angles and lighting improve accuracy.

From a URL:

Tool: mcp__twelvelabs-mcp__create-asset
Parameters:
  url: "<publicly accessible image URL>"

From a local file:

Tool: mcp__twelvelabs-mcp__create-asset
Parameters:
  file: "<absolute path to image file>"

Repeat for each reference image. Collect all returned asset IDs.

Step 4: Create the Entity

Create an entity within the collection, linking it to the reference image assets:

Tool: mcp__twelvelabs-mcp__create-entity
Parameters:
  collectionId: "<collection ID from step 2>"
  name: "<person's name>"
  assetIds: ["<asset_id_1>", "<asset_id_2>"]

The response will include the entity ID and the search query format.

Step 5: Search for the Entity

Use the search tool with the entity ID in the query:

Tool: mcp__twelvelabs-mcp__search
Parameters:
  query: "<@entity_id> action description"

Query format: Wrap the entity ID with <@ and >, then add an optional action description.

Examples:

• <@abc123> - find all appearances
• <@abc123> is giving a presentation - find specific actions
• <@abc123> is walking - find movement scenes

Step 6: Display Results

Format results the same way as regular search:

• Filename of the video
• Stream URL (if available)
• Bulleted list of matching segments with timestamps

Managing Entities

List entities in a collection:

Tool: mcp__twelvelabs-mcp__list-entities
Parameters:
  collectionId: "<collection ID>"

Delete an entity:

Tool: mcp__twelvelabs-mcp__delete-entity
Parameters:
  collectionId: "<collection ID>"
  entityId: "<entity ID>"

Delete a collection:

Tool: mcp__twelvelabs-mcp__delete-entity-collection
Parameters:
  collectionId: "<collection ID>"

Example Interaction

User: "I want to find all clips of Sarah in my videos. Here's her photo: https://example.com/sarah.jpg"

Steps:

1. Create entity collection "People"
2. Create asset from the photo URL
3. Create entity "Sarah" with the asset
4. Search with <@entity_id> query
5. Display matching segments

Tips for Better Results

• Use 3-5 reference images from different angles for best accuracy
• Images should clearly show the person's face
• Separate unrelated groups into different collections
• The visual and audio search options are used for entity search

Important Notes

• Marengo 3.0 Required: Entity search only works with indexes created using Marengo 3.0
• Beta Feature: Entity search is currently in beta
• Plan Limits: Free plan allows 1 collection with up to 15 entities