Skip to main content
The Knowledge Base gives your agent access to a collection of documents. When a user asks a question, the agent retrieves the most relevant passages and uses them to construct an accurate, grounded answer. This technique is called RAG — Retrieval-Augmented Generation.

​
Prerequisites

None — embeddings are handled automatically by the Alquimia runtime. Just open the Knowledge Base step and start uploading.

​
Uploading documents

Drag and drop files into the upload area or click to browse. Uploaded documents are:
  1. Processed and chunked into passages
  2. Embedded by the runtime engine
  3. Stored in the vector database (Qdrant)

​
Retrieval parameters

These parameters control how the agent searches your documents when answering a question.

​
Search type

TypeDescriptionBest for
similarityPure semantic similarity — finds the most relevant chunksSpecific, focused questions
mmrMaximal Marginal Relevance — balances relevance with diversityBroad questions, avoiding repetitive passages

​
Parameters

ParameterDefaultDescription
K4Number of document chunks retrieved per query
Score ThresholdnoneMinimum similarity score (0.0–1.0). Chunks below this are excluded.
Fetch K—MMR only: candidates to evaluate before selecting the final K
Lambda Mult—MMR only: diversity weight (0 = max diversity, 1 = max relevance)
Start with similarity, K=4, and no threshold. Run a few test queries in the Test Drawer. If answers are repetitive, switch to mmr. If answers include irrelevant content, add a score threshold (try 0.7 first).

​
How retrieval works at runtime

When the agent receives a message:
  1. The query is embedded by the runtime using the same model that indexed the documents
  2. The vector database finds the K most similar document chunks
  3. Those chunks are injected into the agent’s prompt as context
  4. The LLM uses this context to generate a grounded answer
The agent cites information from your documents rather than relying solely on what the LLM learned during training.

​
Next steps

Memory

Add conversation memory alongside document retrieval.

Model Selection

Choose the LLM that generates grounded answers from retrieved context.