Knowledge Base

​

Introduction

The Knowledge Base (KB) in Lyzr empowers AI agents to retrieve and utilize both structured and unstructured information for accurate, context-aware responses. It supports various file formats, advanced chunking strategies, and multiple retrieval methods to ensure high-quality information extraction.4

​

Creating and Managing a Knowledge Base

Lyzr provides a streamlined interface via Lyzr Studio to manage Knowledge Bases:

​

Create a Knowledge Base

• Configure embedding, LLM, and vector store credentials.
• Set retrieval and chunking strategies.
• Define a unique name and description.

​

Manage Content

• Add content: Upload documents, enter text, or provide URLs.
• Delete content: Remove outdated or irrelevant entries.
• Update configuration: Change retrieval types or chunk settings anytime.

​

Supported File Types

The following formats can be uploaded to a Lyzr KB:

• .pdf
• .doc, .docx
• .txt
• Website URLs (via scraping)

​

Upload Limitations

To ensure optimal performance:

• Max 5 files at a time
• Each file must be less than 15MB
• For better results, prefer batch-wise uploading

​

Chunking Strategy

Chunking splits documents into smaller parts for better semantic indexing.

​

Parameters:

• Number of chunks: Number of sections generated.
• Chunk size: Controls the length of each chunk.
• Overlap: Adds context continuity across chunks.

This improves both retrieval quality and answer coherence.

​

Available Retrieval Types

Lyzr offers multiple retrieval mechanisms to suit different information needs:

​

a) Basic Retrieval

• Default vector similarity-based retrieval.
• Great for general knowledge lookups.

​

b) MMR (Maximal Marginal Relevance)

• Balances diversity and relevance.
• Reduces duplicate content in retrieved results.

​

c) HyDE (Hypothetical Document Embeddings)

• Generates synthetic documents to simulate context.
• Boosts open-ended query results.

​

Retrieval-Augmented Generation (RAG)

Lyzr seamlessly integrates RAG to generate more accurate and grounded answers using knowledge base content.

​

RAG Workflow

1. Query Reception
   Agent receives a user question or instruction.
2. Document Retrieval
   Top-N relevant documents are fetched using vector similarity.
3. Reranking & Filtering
   Results are optionally refined for relevance.
4. Prompt Assembly
   Retrieved context is combined with the original question.
5. Generation
   LLM generates a grounded response using the assembled prompt.
6. Citation & Delivery
   Output includes references to source documents for transparency.

​

Core Components

• Vector Store: Stores semantic vectors (e.g., Pinecone, FAISS, Qdrant)
• Embedding Model: Transforms content into vectors (e.g., OpenAI, Cohere)
• Reranker: Improves result ordering (optional)
• Prompt Template: Defines how context + question are structured
• Citation Module: Appends references to the output

​

Simulator Testing

Once a Knowledge Base is created and populated:

1. Navigate to the Agent Simulator in Lyzr Studio.
2. Select the agent connected to your KB.
3. Enter test prompts to evaluate:
   • Retrieval accuracy
   • Answer relevance
   • Citation correctness
4. Adjust retrieval type, chunking, or KB content as needed.

Testing helps validate that the agent understands and uses the KB effectively before production deployment.

​

Conclusion

Lyzr’s Knowledge Base system is a robust tool for enabling intelligent, grounded, and flexible AI responses. With support for diverse file types, retrieval strategies, and RAG integration, it provides a powerful foundation for domain-specific agents.

Optimize your AI workflows by:

• Configuring proper chunking
• Choosing the right retrieval type
• Uploading high-quality content in batches
• Testing thoroughly with the simulator