How to Enforce Fine-Grained Authorization in Elasticsearch RAG with SpiceDB

Summary
A practical guide to using SpiceDB relationship-based access control in Elasticsearch RAG pipelines, with a deterministic authorization step that filters retrieved documents before they reach an LLM.
Key points
- The proposed flow retrieves relevant chunks from Elasticsearch, then checks each document ID against SpiceDB before passing authorized content to the LLM.
- Relationship-based access control (ReBAC) represents permissions as relationships between users, groups, and documents, supporting document sharing and individual exceptions without reindexing.
- A deterministic LangGraph authorization node must run before generation; missing document IDs, denied permissions, or individual check errors result in documents being dropped.
- If SpiceDB is unreachable, the pipeline stops before generation, preventing the LLM from receiving unchecked documents.
- Post-filtering can reduce the number of usable results, so the article recommends overfetching and using coarse metadata filters where appropriate.
- Production guidance includes synchronizing relationships with the access-control source of truth and deploying SpiceDB with a persistent datastore.
Article Details
- Defense Focus
- Prevent unauthorized documents from reaching the LLM in Elasticsearch RAG pipelines by enforcing fine-grained, relationship-based authorization outside the model.
- Detection Methods
- Check each retrieved document's doc_id against the user's view permission in SpiceDB before passing it to the LLM.
- Log authorization outcomes, including authorized and denied counts and denied document IDs, to support review of access decisions.
- Fail closed: drop documents with missing doc_id or an individual authorization error, and stop the pipeline if SpiceDB is unreachable.
- Data Sources
- Retrieved document metadata, including doc_id
- SpiceDB relationship and permission-check results
- Authorization logs containing subject_id, document_count, authorized count, denied count, and denied document IDs
- Rule Types
- SpiceDB schema and relationship tuples defining document view permissions
- SpiceDB CheckPermission and CheckBulkPermissions API calls
- Elasticsearch kNN search with an optional metadata terms pre-filter
- Defensive Actions
- Place a deterministic authorization node between retrieval and generation, with no route from retrieval directly to the LLM.
- Pass only documents confirmed as authorized by SpiceDB into the generation prompt; do not expose details about denied documents.
- Keep access relationships synchronized with the source of truth, writing relationships for grants and deleting them for revocations.
- Use overfetching to compensate for documents removed by post-filter authorization, while accounting for Elasticsearch candidate limits.
- Use coarse metadata filters during kNN retrieval when appropriate, while retaining SpiceDB for fine-grained authorization.
- Avoid reindexing when document-sharing relationships change; keep the relationship graph in SpiceDB.
- Use a persistent datastore for production SpiceDB deployments.
Vendors
Products
ElasticsearchElasticsearch retrieves on relevance, and then each chunk is checked against the graph before the large language model (LLM) sees it.MistralThe user's query goes through an embedding model (Mistral, in our repo) to generate a query embedding.SpiceDBYou can implement fine-grained authorization in a relationship graph with SpiceDB.
