Vector Management

Traditional AI systems store vectors in unified indexes, separating them via partition filters. This approach exposes developers to data leaks if bugs occur in partition handling.

Omni uses Enclave Isolation. Each workspace user receives a dedicated vector namespace inside our Cloudflare Vectorize indices.

---

Vector Search Features

• Isolated Vector Spaces: The search query is physically restricted to the namespace belonging to the developer's project ID / user ID.
• Auto Embedding: Text embeddings are automatically generated during facts ingestion using the e5-multilingual-small engine (384 dimensions).
• Client-Side Fallback: Developers can opt-in to calculate embeddings directly on client devices, avoiding plain-text transmission over public networks.

---

Provider Compatibility Notes

Omni's vector layer is designed to support provider selection over time. Provider adapters should preserve the same logical metadata contract even when the underlying vector database has different limits.

• Stable Logical Metadata: Store projectId, workspaceId, documentId, and blockId in vector metadata so search results can be resolved back to PKMS documents.
• Compact Physical IDs: Some providers impose strict record ID limits. Cloudflare Vectorize currently rejects IDs over 64 bytes, so gateway-generated PKMS vector IDs use deterministic short hashes instead of readable workspaceId:documentId:blockId strings.
• Dimension Alignment: The embedding model dimension must match the index dimension. The managed e5-multilingual-small path uses 384 dimensions; any provider swap must provision or select an index with matching dimensions.
• Fallback State: If vector insertion fails but plaintext indexing succeeds, the workspace reports plaintext_indexed and searches continue through full-text fallback until semantic indexing is repaired.