What should I use for permissions-aware retrieval over company files?

For permissions-aware retrieval over company files, the best choice is usually a hybrid architecture:

Recommended approach

1) Use a vector store for semantic search

Store embeddings for:

• documents
• chunks/paragraphs
• metadata like doc_id, department, owner, acl_group_ids, allowed_user_ids, classification, etc.

This gives you good recall for “find the thing I mean.”

2) Enforce permissions with metadata filters

Before returning results, filter by the user’s effective permissions:

• user ID
• group memberships
• roles
• document-level ACLs
• folder-level inheritance
• tenant/org boundaries

This is the most important part: retrieval must be permission-filtered at query time, not only at indexing time.

3) Optionally add a search engine for keyword + hybrid ranking

For many company-file use cases, a combination of:

• BM25 / keyword search
• vector search
• ACL filtering

works better than vector-only.

---

Good technology choices

If you want a managed/vector-first stack

• Pinecone: good metadata filtering and scalable vector search
• Weaviate: strong hybrid search + filters
• Qdrant: excellent payload filtering and straightforward ACL metadata use
• Milvus: scalable, but ACL/filtering ergonomics vary depending on setup

If you want classic enterprise search

• Elasticsearch / OpenSearch
  • great for hybrid search
  • strong metadata filtering
  • can support permissions-aware retrieval well
  • often a good fit if your data is mostly documents and keyword search matters a lot

If you need strong access-control integration

Consider using your source-of-truth authorization system:

• AWS IAM / Azure AD / Okta / Google Workspace groups
• document ACLs from SharePoint, Drive, Box, S3, etc.
• enterprise policy engines like OPA (Open Policy Agent) or Cedar

Then pass the allowed IDs into the retrieval layer as filters.

---

What usually works best in practice

For company files, I’d recommend:

Embeddings + metadata ACL filtering + hybrid search

• Index chunks with embeddings
• Attach permission metadata to every chunk
• At query time, compute the user’s authorized scope
• Filter results before ranking/final answer generation

This avoids leakage and still gives good search quality.

---

Important caveat

Don’t rely on the LLM to “respect permissions” by prompt alone.
Permissions must be enforced in the retrieval layer and/or backend authorization layer.

---

If you want a simple default recommendation

• Qdrant or Pinecone for vector retrieval
• Elasticsearch/OpenSearch if keyword/hybrid search is critical
• OPA/Cedar or your identity provider for permissions decisions
• Store ACL/group metadata on every chunk and filter at query time

If you want, I can also suggest a concrete architecture for:

1. Google Drive/SharePoint-style file permissions, or
2. an implementation in Pinecone/Qdrant/OpenSearch.