🤖 Explicar con IA

Knowledge Base (RAG)¶

Give your AI agents domain-specific knowledge by uploading documents. The knowledge base uses Retrieval-Augmented Generation (RAG) to provide accurate, context-aware responses grounded in your own data.

Overview¶

The knowledge base system:

1. Upload — Add documents (PDF, DOCX, TXT, CSV)
2. Process — Documents are chunked and embedded
3. Store — Embeddings are stored for semantic search
4. Retrieve — When a user asks a question, relevant chunks are retrieved
5. Generate — The AI uses retrieved context to generate accurate responses

Uploading Documents¶

Via the Agent Studio¶

1. Open the Agent Studio (agent creation wizard)
2. Navigate to the Knowledge Base step
3. Click Upload Documents
4. Select files from your computer
5. Wait for processing to complete
6. Click Continue

Supported File Types¶

Processing¶

When a document is uploaded:

1. Text Extraction — Content is extracted from the file
2. Chunking — Text is split into semantic chunks (~500 tokens each)
3. Embedding — Each chunk is converted to a vector embedding
4. Storage — Embeddings are stored in the knowledge base

Processing typically takes 1-5 minutes depending on document size.

Managing Knowledge Files¶

Viewing Uploaded Files¶

1. Open the agent details
2. Navigate to the Knowledge Base section
3. View the list of uploaded files:

Adding More Documents¶

You can add documents at any time:

1. Open the agent details
2. Click Add Documents
3. Upload new files
4. Wait for processing

New documents are immediately available for retrieval.

Removing Documents¶

1. Open the knowledge base
2. Find the file to remove
3. Click Delete
4. Confirm removal

The file’s chunks and embeddings are permanently deleted.

How RAG Works¶

Retrieval Process¶

When a user sends a message to the agent:

graph LR
    A[User Message] --> B[Embed Query]
    B --> C[Search Knowledge Base]
    C --> D[Retrieve Top Chunks]
    D --> E[Inject Context]
    E --> F[Generate Response]

1. Embed Query — The user’s message is converted to a vector
2. Semantic Search — The vector is compared against stored embeddings
3. Retrieve Chunks — The most relevant chunks are selected (top-k)
4. Inject Context — Retrieved chunks are added to the AI’s prompt
5. Generate Response — The AI generates a response using the context

Retrieval Configuration¶

Best Practices for Knowledge Bases¶

Document Quality¶

• Use clear, well-structured documents — The AI can only retrieve what’s there
• Include headers and sections — Helps with chunking quality
• Keep documents current — Remove outdated information
• Be specific — Detailed documents produce better retrieval results

Document Organization¶

• One topic per document — Avoid mixing unrelated topics
• Use consistent formatting — Standardize document structure
• Name files descriptively — Makes management easier
• Version control — Re-upload when documents change significantly

Common Use Cases¶

Storage¶

Knowledge base files are stored in the agent-knowledge Supabase Storage bucket:

• Files are stored per-user in isolated folders
• Access is controlled by authentication
• Files are encrypted at rest
• Processing is handled by the process-knowledge edge function

Limitations¶

Troubleshooting¶

Document stuck in “Processing”¶

• Wait up to 10 minutes for large files
• Check file format is supported
• Try re-uploading the file
• Contact support if the issue persists

Agent not using knowledge base¶

• Verify documents are in “Ready” status
• Check that the query is related to the document content
• Try rephrasing the question
• Add more relevant documents to the knowledge base

Poor response quality¶

• Review document quality and structure
• Add more specific documents
• Check retrieval settings (top-k, threshold)
• Ensure documents are up to date

What’s Next¶

• Website Analysis — Auto-generate prompts from websites
• Voice AI — Set up voice agents with phone integration

Aún con dudas? Pregunta en Discord o explore tutoriales