An AI analyst needs permission-aware search because internal emails, meeting notes, reports, and research documents can each have different access rules.
2
Post-filtering loses relevant results, while pre-filtering can require scanning tens of millions of documents, so neither approach works well at large scale.
3
turbopuffer combines permission indexes with clustered vector search and reports 120 millisecond P90 searches across 400 million vectors.
Summary
Aaron Linsky explains why Bridgewater's AI analyst needs access to private research while respecting document-level permissions. Bridgewater's Vault contains about 100 million documents from external providers and internal sources, and its first permission approach failed as the collection grew. Nikhil Benesch then explains the search problem. Scanning every vector is too slow, post-filtering can miss documents a user is allowed to see, and pre-filtering can leave too many documents to scan. turbopuffer's native filtering combines both methods. It identifies vector clusters containing documents the user may access, searches the clusters nearest to the query, and scans within those smaller groups. Benesch says this produces 120 millisecond P90 searches across 400 million vectors. The talk also shows how Bridgewater maps user and group permissions into filter expressions, while noting that nested groups remain difficult. turbopuffer is extending the same approach to regional, VPC, and customer-controlled encryption isolation.
An AI analyst must search private data under document-level permissions
The analyst starts with a question, asks an LLM to investigate, and retrieves facts from public and private sources. Public documents have the same visibility for everyone in an organization. Internal emails, meeting notes, and reports can each belong to different groups or individual users. That means the search system must apply permissions while finding relevant evidence. Nikhil Benesch says this is a classic information-retrieval problem and one that turbopuffer was designed to address.
Bridgewater's research process is moving from human-led work toward machine-led work
Aaron Linsky describes Bridgewater's process as a five-step research circle: perception, hypothesis, investigation, systematization, and industrialization. The firm has encoded market and economic understanding in systems over decades, moving from a bond system written on a yellow legal pad in 1980, through Excel and Scala, to Python. Bridgewater now wants machines to take a larger role in this process, which requires access to both time-series data and high-quality unstructured research.
The Vault combines external research, internal material, and permissions at large scale
Bridgewater's Vault contains about 100 million documents. The collection includes research from 135 external sources, internal writing, meeting recordings, and internal reports. Some external sources are unavailable to everyone at Bridgewater, so the documents carry different permissions. Investors and AI agents read from the Vault, then create new material that goes back into it. Linsky says the first implementation seemed to work until the document collection grew, at which point it failed badly.
Benesch uses simple throughput arithmetic to explain why scanning every document fails. If each document occupies one kilobyte and memory can be read at about 20 gigabytes per second, scanning 100 million documents takes about five seconds, even with all data already in memory. Search users generally want results in about 100 milliseconds. If the data is still on object storage, loading it can take a minute or more. A vector index is needed to avoid examining the entire collection.
A vector index finds similar concepts by searching nearby vectors
Vector embeddings turn concepts into numeric coordinates. Benesch uses red shoes, maroon sandals, and a concrete block to illustrate that similar concepts should be near one another in vector space. In a simplified two-dimensional picture, a query for red shoes is plotted on a graph and the nearby vectors are searched. This lets the system inspect a much smaller group, rather than all documents. The same idea applies when the vectors have hundreds or thousands of dimensions.
Permissions make an ordinary vector index return the wrong candidate documents
The documents nearest to a query are not necessarily visible to the user making that query. Benesch describes post-filtering as retrieving the closest 10,000 vectors and then discarding documents that fail the permission check. This can miss the best visible result, because the system may stop before it reaches documents the user can read. A user searching for red shoes might know that a maroon-sandals document exists, but post-filtering may never return it.
Pre-filtering preserves permissions but can discard the speed advantage of the vector index
The alternative is to start with the permission filter, find every document the user can access, and then compare those documents with the query vector. That works logically, but a user who can access 50 million of the Vault's 100 million documents still leaves the system with a huge sequential scan. Benesch says this returns to the original performance problem, because the vector index is no longer narrowing the search enough.
Native filtering combines permission-aware clusters with local vector scans
turbopuffer clusters similar vectors and records which permission sets have access to documents in each cluster. For a user such as Aaron, the system can discard clusters containing no accessible documents, search the nearest remaining clusters, and scan within those clusters. Benesch says the clusters currently contain about 200 documents and that the size is still being tuned. This combines pre-filtering and post-filtering into what turbopuffer calls native filtering.
The reported result is 120 millisecond P90 search across hundreds of millions of vectors
Benesch says the system reaches 120 millisecond P90 searches across 450 million vectors, with Bridgewater's 100 million documents becoming 400 million vectors. Bridgewater supplies its permission model as filter expressions such as viewers containing a user or groups containing any of two groups. turbopuffer handles the query planning needed to apply those logical filters at scale. Benesch also says nested groups remain difficult and require further engineering.
"Every email, every meeting note, every report might have a completely different set of permissions attached to it, belonging to different groups, maybe individual permissions assignments on those docs."Nikhil Benesch01:27
Who should watch
You are building retrieval-augmented generation over documents with per-user or per-group access rules.
Your vector search is fast without filters but loses recall or latency once permissions are applied.
You need to think about regional, VPC, or customer-controlled encryption isolation alongside document permissions.