Store embeddings, metadata, and document references in Milvus; store raw documents and large files externally to minimize collection size.
Milvus storage:
1. Vector embeddings: Dense embeddings (768–1536 dims) from OpenAI, Anthropic, or open-source models. This is the core retrieval data.
2. Metadata fields:
- Document ID, source, ingestion timestamp
- Document type (contract, invoice, email, code)
- Author, department, classification level (security sensitive)
- Date ranges, version numbers
- Agent accessibility flags (“compliance_agent=true”)
3. Sparse embeddings: BM25/keyword vectors for hybrid search. Store alongside dense embeddings.
4. Summary tokens: If documents are chunked, store chunk summaries (50–100 tokens) in a text field. Agents use summaries to decide relevance.
External storage (S3, GCS):
- Full document text
- Images, PDFs, binary files
- Large attachments
- Historical versions
Milvus retrieves document IDs + metadata; agents fetch full content from external storage if needed.
Data example: json { "id": "doc_12345", "embedding": [0.23, -0.15, …], "sparse_embedding": {"invoice": 0.8, "payment": 0.6}, "metadata": { "source": "s3://bucket/invoices/INV-2025-00142", "doc_type": "invoice", "date": 1700000000, "agent_filter": ["finance", “audit”] }, "summary": “Invoice for 500 units of SKU-X, $50k, Q3 2025” }
This design keeps Milvus fast and scalable while agents access rich context.
Related Resources: