The same corpus. Pointed at your problem.
haystak is one engine, but the shape of the problem changes by team. Here's where it fits first, each built on the same permission-trimmed corpus.
Put AI on your content without opening the vault.
Teams want RAG and agents over internal content, but can't risk surfacing what a user shouldn't see. haystak trims every AI call to the caller's grants, fail-closed, and RAG answers cite the source records they drew from.
- Every call permission-trimmed to the caller
- Cited answers: open the source and check
- Runs on your own models, on your hardware
- Files and databases in one result set
- Schema-aware joins across sources
- Row-level ACL on every result
One search box over shares and systems of record.
Content is split across file shares and databases with no single view. haystak returns cross-source results with schema-aware joins in one permission-trimmed result set, for people, agents, or analysts.
Your analysts' SQL tools, pointed at the corpus.
Documents and database records are usually queried in different places. haystak presents the whole corpus over the SQL Server wire protocol, so Power BI, Excel, and SSMS read files and records as ordinary tables, where they already are. Nothing is copied into a warehouse first, and every row is trimmed to the person connected.
- Files and records as tables
- Nothing moved into a warehouse
- Permission-trimmed at the wire
AI on regulated content, nothing leaving the wire.
Sensitive environments can't ship data to a cloud. haystak indexes in place and runs local models (Ollama or any OpenAI-compatible server), so the whole pipeline can run fully air-gapped, with the same permissions and audit.
- Index in place: no upload, no egress
- Local inference on your hardware
- Air-gapped end to end
More places the corpus earns its keep.
Data & content discovery
Find what you already have across silos, with provenance on every value: where a fact came from, and what's duplicated.
Matter administration
Open a matter, place it on hold, close it, and dispose of it as a whole, with every read and change on the record.
Regulated & permission-sensitive work
Read-only sources, tamper-evident audit, deletion-as-suppression, directory identity across every surface, and search completeness validation (a reproducible, court-recognized method for proving a search didn't miss what it should have found).
Which problem is yours?
A focused walkthrough on a corpus that looks like yours, scoped to the use case you care about.