Knowledge bases

A knowledge base in Meko is a RAG (Retrieval-Augmented Generation) pipeline that ingests your documents, breaks them into chunks, generates vector embeddings, and indexes them for semantic search. Agents can then query the knowledge base to find relevant information from your documents.

How it works

When you add a knowledge base to a datapack, Meko's pg_dist_rag pipeline:

1. Fetches documents from the source (S3, local filesystem, or NFS).
2. Preprocesses using unstructured to extract text from PDFs, images, parquet, iceberg, JSON, and more.
3. Chunks the text into segments (configurable chunk size).
4. Embeds each chunk using the configured embedding model.
5. Indexes the embeddings in pgvector for fast similarity search.

All of this happens within your datapack's database; there's no separate vector database to manage.

Supported document formats

Meko supports documents in:

• PDF
• Parquet
• Iceberg
• JSON
• Images
• Video

Documents can be loaded from S3, local filesystem, or NFS.

Add a knowledge base

Using the CLI:

meko datapack add_knowledge_base --name my-datapack --url s3://bucket/docs/

You can configure chunking settings:

meko datapack add_knowledge_base --name my-datapack \
  --url s3://bucket/docs/ \
  --settings '{"chunk_size": 512}'

Using the API:

POST /api/v1/datapacks/{name}/knowledge-bases

{
  "url": "s3://bucket/docs/"
}

Query knowledge

Once indexed, agents can query the knowledge base through the MCP server. The MCP tools handle embedding the query, performing similarity search, and returning relevant chunks.

Manage knowledge bases

Next steps

• Work with knowledge bases - How to build and query knowledge bases
• Learn about datapacks