Creating a Knowledge Base
- Go to Knowledge Base and select + New.
- Select Basic as the Knowledge Base Type. This option uses simple vector-based retrieval with embeddings.
- Enter a Name (letters, numbers, and underscores only) and an optional Description.
- Select a Vector Store and LLM Embedding Model.
- Select Create Knowledge Base.
- Add content through file upload, text, URL, or live source.
- Train the KB.
- Attach it to an agent through the agent builder’s Knowledge Base feature.
Supported file types
.pdf.doc.docx.txt- Website URLs
Upload limitations
| Limit | Value |
|---|---|
| Files per upload | 5 |
| File size | Less than 15 MB each |
| Recommendation | Upload in batches and test retrieval quality between batches |
Chunking strategy
Chunking controls how documents are split before embedding. Smaller chunks improve precision; larger chunks preserve context.| Setting | Description |
|---|---|
| Chunk size | Maximum number of tokens in each chunk |
| Overlap | Number of tokens shared between adjacent chunks, preserving context at boundaries |
| Number of chunks | Maximum number of chunks returned per query |
Retrieval types
| Type | Best for |
|---|---|
| Basic Retrieval | General vector similarity search |
| MMR (Maximal Marginal Relevance) | Reducing duplicate chunks while preserving relevance |
| HyDE (Hypothetical Document Embeddings) | Improving retrieval accuracy on open-ended or vague queries |
Retrieval modes
Retrieval mode controls how the agent uses the Knowledge Base to answer a query. It is a separate choice from the retrieval type above, which controls the similarity algorithm. Three modes are available:| Mode | How it works | Best for |
|---|---|---|
| Basic RAG (default) | Retrieves relevant chunks and generates a response from them in a single pass. | General question answering. |
| Agentic RAG | Runs one or more retrieval cycles and reasons across the results before answering. | Questions that need iterative reasoning or several retrievals. |
| One-Shot RAG | Retrieves relevant information in a single retrieval step and generates the response in one LLM call. | Straightforward question answering where a single retrieval provides enough context, without iterative reasoning or multiple retrieval cycles. |
Score threshold
The score threshold filters out chunks whose similarity score falls below a minimum value. Raising the threshold improves answer precision but may reduce recall on borderline queries. Start at the default and adjust based on test results.Playground Retrieval
Once the KB is trained, open it and use the Playground Retrieval panel on the right to test retrieval before attaching it to an agent. Type a query in the input field and select Retrieve to see the matching chunks and their similarity scores. This lets you verify that the right content is surfacing for representative questions and catch chunking or configuration issues before doing a full agent deployment.
Live Sources
Live Sources automatically sync content on a configurable frequency. When new content is detected, Lyzr adds the delta instead of re-ingesting everything. Available for:- SharePoint: syncs documents from selected SharePoint sites
- Website: crawls and re-indexes updated pages
- Google Drive: planned support