Configuring Retrievers

Retrievers are configurations for RAG (Retrieval-Augmented Generation) functionality in Wegent. They define how documents are indexed, stored, and retrieved.

Prerequisites

• Wegent platform installed and running
• Elasticsearch service enabled (optional, only needed for RAG features)

  docker compose --profile rag up -d

What is a Retriever?

A Retriever is a CRD (Custom Resource Definition) that configures:

• Storage Backend: Vector database connection (Elasticsearch, Qdrant)
• Index Strategy: How documents are organized in the database
• Retrieval Methods: Search modes (vector, keyword, hybrid)
• Embedding Configuration: How text is converted to vectors

Creating a Retriever

Via Web UI

1. Navigate to Settings → Retrievers
2. Click Add Retriever
3. Fill in the configuration:
   • Name: Unique identifier (e.g., my-es-retriever)
   • Display Name: Human-readable name
   • Storage Type: Select elasticsearch or qdrant
   • URL: Storage backend URL (e.g., http://elasticsearch:9200)
   • Authentication: Username/password or API key (optional)
   • Index Strategy: Choose indexing mode
   • Retrieval Methods: Enable vector, keyword, or hybrid search
4. Click Test Connection to verify settings
5. Click Create to save

Via API

POST /api/retrievers
Content-Type: application/json

{
  "apiVersion": "agent.wecode.io/v1",
  "kind": "Retriever",
  "metadata": {
    "name": "my-es-retriever",
    "namespace": "default",
    "displayName": "My Elasticsearch Retriever"
  },
  "spec": {
    "storageConfig": {
      "type": "elasticsearch",
      "url": "http://elasticsearch:9200",
      "username": "elastic",
      "password": "password",
      "indexStrategy": {
        "mode": "per_user",
        "prefix": "wegent"
      }
    },
    "retrievalMethods": {
      "vector": {
        "enabled": true,
        "defaultWeight": 0.7
      },
      "keyword": {
        "enabled": true,
        "defaultWeight": 0.3
      },
      "hybrid": {
        "enabled": true
      }
    },
    "description": "Elasticsearch retriever for RAG"
  }
}

Index Strategies

Choose an index strategy based on your use case:

Strategy	Description	Best For
per_user	One index per user	Elasticsearch deployments, user-level isolation
per_dataset	One index per knowledge base	Multi-tenant scenarios, dataset isolation
fixed	Single fixed index	Small datasets, simple setup
rolling	Hash-based sharding	Large datasets, load distribution

Recommended: per_user Mode

For Elasticsearch, we recommend using per_user mode:

{
  "indexStrategy": {
    "mode": "per_user",
    "prefix": "wegent"
  }
}

This creates indices like wegent_user_123, providing better performance and isolation.

Retrieval Methods

Vector Search (Semantic)

Pure vector similarity search for semantic understanding:

Parameter	Description	Default
retrieval_mode	Retrieval mode	vector
top_k	Number of results	5
score_threshold	Relevance threshold	0.5

Use cases: Concept matching, understanding questions, semantic search

Keyword Search

Traditional BM25 keyword matching:

Parameter	Description	Default
retrieval_mode	Retrieval mode	keyword
top_k	Number of results	5

Use cases: Exact term matching, code search, API names

Hybrid Search (Vector + Keyword)

Combines vector similarity with BM25 keyword matching:

Parameter	Description	Default
retrieval_mode	Retrieval mode	hybrid
vector_weight	Vector weight	0.7
keyword_weight	Keyword weight	0.3
top_k	Number of results	5
score_threshold	Relevance threshold	0.5

Weight recommendations:

• Conceptual queries (0.8/0.2): Understanding, explanations
• Balanced (0.7/0.3): General purpose (default)
• Precise matching (0.3/0.7): Code search, API names, exact terms

---

Retrieval Test

Before saving retrieval configuration, you can use the retrieval test feature to verify effectiveness.

How to Use

1. Go to Knowledge Base Retrieval Settings
2. Configure retrieval parameters (mode, top_k, threshold, etc.)
3. Enter a test query in the Retrieval Test area
4. Click Test button
5. Review returned document chunks and relevance scores
6. Adjust parameters based on results
7. Click Save when satisfied

Test Recommendations

Test Type	Suggested Query
Conceptual	Use descriptive questions like "What is..."
Precise	Use specific terms or names
Boundary	Use vague or unrelated queries

Retrieval test helps you optimize retrieval configuration before actual use.

Using Retrievers with RAG

1. Retrieve Relevant Chunks

POST /api/rag/retrieve
Content-Type: application/json

{
  "query": "How do I configure a bot?",
  "knowledge_id": "kb_001",
  "retriever_ref": {
    "name": "my-es-retriever",
    "namespace": "default"
  },
  "embedding_config": {
    "provider": "openai",
    "model": "text-embedding-3-small",
    "api_key": "sk-..."
  },
  "top_k": 5,
  "score_threshold": 0.5,
  "retrieval_mode": "hybrid",
  "hybrid_weights": {
    "vector_weight": 0.7,
    "keyword_weight": 0.3
  }
}

Embedding Providers

OpenAI

{
  "provider": "openai",
  "model": "text-embedding-3-small",
  "api_key": "sk-...",
  "base_url": "https://api.openai.com/v1"  // Optional
}

Custom API (OpenAI-compatible)

{
  "provider": "custom",
  "model": "your-model-name",
  "api_key": "your-api-key",
  "base_url": "https://your-api-endpoint.com/v1"
}

Resource Scopes

Retrievers support three scopes:

Scope	Description	Access
Personal	Your private retrievers	Only you
Group	Shared within a group	Group members
Public	System-provided retrievers	All users

Best Practices

1. Index Strategy Selection

• Use per_user for Elasticsearch (recommended)
• Use per_dataset for multi-tenant scenarios with dataset isolation
• Avoid fixed for production (only suitable for small, single-tenant deployments)

2. Retrieval Mode Selection

• Vector mode: Semantic understanding, concept matching
• Hybrid mode: Balanced semantic and exact matching (recommended for most use cases)
• Adjust weights: Based on query type (conceptual vs. precise)

3. Security

• Store credentials securely (use environment variables or secrets management)
• Use API keys instead of username/password when possible
• Restrict access using namespaces and groups

4. Performance

• Choose appropriate index strategy based on dataset size
• Monitor storage backend performance
• Use per_user mode for Elasticsearch to avoid index explosion
• Set appropriate top_k and score_threshold values

5. Document Management

• Use meaningful knowledge_id values for organization
• Regularly clean up unused documents
• Monitor storage usage

Troubleshooting

Connection Failed

Problem: Cannot connect to Elasticsearch

Solutions:

1. Verify Elasticsearch is running: docker ps | grep elasticsearch
2. Check URL is correct: http://elasticsearch:9200 (internal) or http://localhost:9200 (external)
3. Test connection: Use the Test Connection button in the UI
4. Check credentials: Verify username/password or API key

Indexing Failed

Problem: Document upload fails

Solutions:

1. Check file format is supported (MD, PDF, TXT, DOCX, code files)
2. Verify embedding provider credentials
3. Check Elasticsearch storage capacity
4. Review backend logs for detailed error messages

Low Retrieval Quality

Problem: Retrieved chunks are not relevant

Solutions:

1. Try hybrid mode instead of pure vector mode
2. Adjust hybrid weights based on query type
3. Lower score_threshold to get more results
4. Use a better embedding model (e.g., text-embedding-3-large)
5. Improve document chunking (automatic semantic chunking is used)

API Reference

For complete API documentation, see:

• Backend API docs: http://localhost:8000/api/docs
• AGENTS.md: RAG Services section

Using Knowledge Base Without RAG (No Retriever Mode)

You can create and use knowledge bases even without configuring a retriever. In this mode, the AI uses exploration tools instead of semantic search.

What Works Without RAG

• ✅ Document upload and storage
• ✅ Document viewing and editing
• ✅ AI can browse documents using kb_ls (list) and kb_head (read) tools
• ✅ Manual document exploration by AI
• ✅ Knowledge base chat in notebook mode

What Requires RAG Configuration

• ❌ Semantic search (knowledge_base_search tool)
• ❌ Vector similarity retrieval
• ❌ Automatic chunk-based retrieval
• ❌ Hybrid search (vector + keyword)

When to Use No-RAG Mode

Consider using knowledge bases without RAG when:

1. No Vector Database Available: You don't have Elasticsearch or other vector database set up
2. Small Knowledge Base: Your knowledge base is small enough for AI to read through documents
3. Testing: You want to test knowledge base functionality without RAG infrastructure
4. Cost Optimization: You want to avoid embedding model API costs

AI Behavior in No-RAG Mode

When you chat with an AI that has access to a knowledge base without RAG:

1. Document Discovery: AI uses kb_ls to list available documents with summaries
2. Content Selection: AI reviews document summaries to identify relevant ones
3. Content Reading: AI uses kb_head to read document content (with pagination for large files)
4. Answer Generation: AI answers based on the content it has read

Example Workflow

User: What does the API documentation say about authentication?

AI: Let me explore the knowledge base to find relevant information.

[Uses kb_ls to list documents]
Found 5 documents:
- api-guide.md (15KB) - API usage guide with authentication section
- setup.md (8KB) - Initial setup instructions
- ...

[Uses kb_head to read api-guide.md]
Reading authentication section from api-guide.md...

Based on the API documentation, authentication uses JWT tokens...

Performance Considerations

This approach is less efficient than RAG retrieval:

Aspect	RAG Mode	No-RAG Mode
Search Speed	Fast (vector similarity)	Slower (sequential reading)
Token Usage	Lower (relevant chunks only)	Higher (may read full documents)
Accuracy	Semantic understanding	Depends on document summaries
Best For	Large knowledge bases	Small knowledge bases (<50 docs)

Choosing The RAG Mode On Creation

When creating a knowledge base, RAG Retrieval in Advanced Settings provides two modes:

Mode	Behavior	Best For
Auto	The system selects an available retriever and embedding model while preserving retrieval parameters you changed	Recommended default
No RAG	Creates a no-RAG knowledge base without writing retrieval configuration	No vector database, small knowledge bases, or testing

If you choose automatic configuration but the current environment has no usable default retriever or embedding model, the knowledge base can still be created, but no RAG retrieval configuration is written. Uploaded documents are stored, and AI will use knowledge exploration tools to read the content in later conversations.

To create a no-RAG knowledge base:

1. Create Knowledge Base: Open advanced settings in the create dialog and set RAG Retrieval to No RAG
2. Upload Documents: Documents are stored but not indexed for RAG
3. Start Chatting: AI will automatically use exploration tools

Note: After configuring a retriever and embedding model, newly created knowledge bases will automatically receive RAG retrieval configuration.

Related Documentation

• Knowledge Base Guide
• Agent Settings
• Managing Tasks
• YAML Specification

Feedback

Since RAG functionality is experimental, we welcome your feedback:

• Report issues on GitHub
• Suggest improvements
• Share your use cases

---

Note: This feature is under active development. Check the changelog for updates.