Skip to content

Latest commit

 

History

History
360 lines (287 loc) · 9.47 KB

BATCH_OPERATIONS.md

File metadata and controls

360 lines (287 loc) · 9.47 KB

Batch Operations Guide

Home > Docs > Usage > Batch Operations

This guide explains how to efficiently process multiple messages using EverMemOS's batch operations.

Table of Contents

Overview

EverMemOS supports batch processing for efficiently storing multiple messages at once. This is particularly useful for:

  • Processing historical conversation data
  • Importing chat logs from other platforms
  • Group chat conversations with multiple participants
  • Bulk data migration

Group Chat Format

EverMemOS uses a standardized GroupChatFormat for batch operations. This format supports:

  • Conversation metadata (group info, user details)
  • Multi-speaker conversations
  • Timestamps and message IDs

For complete format specifications, see Group Chat Format Specification.

Batch Storage Script

Basic Usage

# Store group chat messages (Chinese data)
uv run python src/bootstrap.py src/run_memorize.py \
  --input data/group_chat_zh.json \
  --api-url http://localhost:1995/api/v1/memories \
  --scene group_chat

# Store group chat messages (English data)
uv run python src/bootstrap.py src/run_memorize.py \
  --input data/group_chat_en.json \
  --api-url http://localhost:1995/api/v1/memories \
  --scene group_chat

# Validate file format without storing
uv run python src/bootstrap.py src/run_memorize.py \
  --input data/group_chat_en.json \
  --scene group_chat \
  --validate-only

Script Parameters

Parameter Required Description
--input Yes Path to the conversation data file (JSON format)
--api-url No API endpoint (default: http://localhost:1995/api/v1/memories)
--scene Yes Scene type: assistant or group_chat
--validate-only No Validate format without sending to API

Scene Parameter Explanation

The --scene parameter specifies the memory extraction strategy:

  • assistant - Use for one-on-one conversations with AI assistant
  • group_chat - Use for multi-person group discussions

Important Note: In your data files, you may see scene values like work, company, or social - these are internal scene descriptors in the data format. The --scene command-line parameter uses different values (assistant/group_chat) to specify which extraction pipeline to apply.

Data Format Specification

GroupChatFormat Structure

{
  "version": "1.0.0",
  "conversation_meta": {
    "group_id": "group_001",
    "name": "Project Discussion Group",
    "description": "Team project planning and updates",
    "scene": "group_chat",
    "timezone": "Asia/Shanghai",
    "user_details": {
      "user_101": {
        "full_name": "Alice",
        "role": "Product Manager",
        "nickname": "Ali"
      },
      "user_102": {
        "full_name": "Bob",
        "role": "Engineer"
      }
    }
  },
  "conversation_list": [
    {
      "message_id": "msg_001",
      "create_time": "2025-02-01T10:00:00+00:00",
      "sender": "user_101",
      "content": "Good morning everyone, let's discuss the new feature"
    },
    {
      "message_id": "msg_002",
      "create_time": "2025-02-01T10:05:00+00:00",
      "sender": "user_102",
      "content": "Sure! I've prepared the technical spec"
    }
  ]
}

Required Fields

conversation_meta:

  • group_id (string) - Unique identifier for the conversation group
  • name (string) - Human-readable name for the group
  • user_details (object) - Map of user IDs to user information

conversation_list:

  • message_id (string) - Unique identifier for each message
  • create_time (string) - ISO 8601 timestamp with timezone
  • sender (string) - User ID (must exist in user_details)
  • content (string) - Message content

Optional Fields

conversation_meta:

  • description (string) - Group description
  • scene (string) - Internal scene descriptor (group_chat or assistant)
  • timezone (string) - Timezone for the conversation

conversation_list:

  • sender_name (string) - Override sender's display name

Examples

Example 1: Simple Group Chat

{
  "version": "1.0.0",
  "conversation_meta": {
    "group_id": "team_standup",
    "name": "Daily Standup",
    "user_details": {
      "alice": {"full_name": "Alice Smith"},
      "bob": {"full_name": "Bob Jones"}
    }
  },
  "conversation_list": [
    {
      "message_id": "msg_1",
      "create_time": "2025-02-01T09:00:00+00:00",
      "sender": "alice",
      "content": "Yesterday I completed the login feature"
    },
    {
      "message_id": "msg_2",
      "create_time": "2025-02-01T09:01:00+00:00",
      "sender": "bob",
      "content": "Great! I'm working on the dashboard today"
    }
  ]
}

###Example 2: Family Chat with Rich Metadata

{
  "version": "1.0.0",
  "conversation_meta": {
    "group_id": "family_chat_001",
    "name": "Smith Family",
    "description": "Family group chat",
    "scene": "group_chat",
    "timezone": "America/New_York",
    "user_details": {
      "mom": {
        "full_name": "Jane Smith",
        "nickname": "Mom",
        "role": "Parent"
      },
      "dad": {
        "full_name": "John Smith",
        "nickname": "Dad",
        "role": "Parent"
      },
      "daughter": {
        "full_name": "Emily Smith",
        "age": 16
      }
    }
  },
  "conversation_list": [
    {
      "message_id": "fam_001",
      "create_time": "2025-02-01T18:00:00-05:00",
      "sender": "mom",
      "content": "Dinner is ready! Come down please.",
    },
    {
      "message_id": "fam_002",
      "create_time": "2025-02-01T18:02:00-05:00",
      "sender": "daughter",
      "content": "Coming! Just finishing homework."
    }
  ]
}

Example 3: One-on-One Assistant Chat

{
  "version": "1.0.0",
  "conversation_meta": {
    "group_id": "user_assistant_001",
    "name": "Personal Assistant",
    "scene": "assistant",
    "user_details": {
      "user_001": {
        "full_name": "Alex"
      }
    }
  },
  "conversation_list": [
    {
      "message_id": "chat_001",
      "create_time": "2025-02-01T10:00:00+00:00",
      "sender": "user_001",
      "content": "I love playing soccer on weekends"
    },
    {
      "message_id": "chat_002",
      "create_time": "2025-02-01T10:30:00+00:00",
      "sender": "user_001",
      "content": "My favorite team is Barcelona"
    }
  ]
}

Command for assistant chat:

uv run python src/bootstrap.py src/run_memorize.py \
  --input my_assistant_chat.json \
  --scene assistant

Best Practices

1. Data Preparation

  • Validate before importing: Use --validate-only to check format
  • Use consistent IDs: Ensure message_id and user IDs are unique
  • Include timestamps: Always use ISO 8601 format with timezone
  • Provide user details: Include at least full_name for each user

2. Performance Optimization

  • Batch size: Process 100-1000 messages at a time for optimal performance
  • Sequential processing: Script processes messages sequentially to maintain order
  • Monitor progress: Watch for errors in terminal output
  • Wait for indexing: Allow 10-15 seconds after completion for search indexes to update

3. Data Quality

  • Clean content: Remove formatting artifacts or special characters
  • Accurate timestamps: Ensure chronological order
  • Complete metadata: Fill in all available user information
  • Meaningful group IDs: Use descriptive, stable identifiers

4. Scene Selection

  • Use assistant for:

    • One-on-one conversations
    • Personal AI assistant chats
    • Individual user interactions

  • Use group_chat for:

    • Multi-participant discussions
    • Team conversations
    • Family or social group chats

Troubleshooting

Validation Errors

Problem: --validate-only reports format errors

Solutions:

  • Check JSON syntax is valid
  • Verify all required fields are present
  • Ensure timestamps are in ISO 8601 format
  • Confirm sender IDs exist in user_details

API Errors

Problem: Script reports API errors when storing

Solutions:

  • Verify API server is running: curl http://localhost:1995/health
  • Check API URL is correct (default: http://localhost:1995/api/v1/memories)
  • Ensure .env has required API keys (LLM_API_KEY, VECTORIZE_API_KEY)
  • Review error messages for specific issues

Slow Processing

Problem: Batch processing is very slow

Solutions:

  • This is normal for large batches (each message requires LLM extraction)
  • Reduce batch size if memory issues occur
  • Ensure Docker services have adequate resources
  • Check LLM API rate limits

Missing Memories

Problem: Messages processed but not searchable

Solutions:

  • Wait 10-15 seconds for indexing to complete
  • Verify Elasticsearch and Milvus are running
  • Check MongoDB for stored data
  • Ensure embeddings were created (requires VECTORIZE_API_KEY)

See Also