Layered grants
Site groups, nested Entra groups, direct grants, sharing links.
DocTalk · a bi concepts service
Talk to your documents.
Safely.
Permission-aware AI search on Snowflake Cortex: every answer filtered by who is asking, with SharePoint permissions enforced inside the index.
Snowflake Cortex SearchSharePoint OnlineMicrosoft Entra ID
01 · The problem
AI search doesn't know who is allowed to read what
AI search runs on an index: every document read once, made findable in milliseconds. The standard build: one pool, for everyone.
chloe.eng · Engineering
“What's the salary band for a Senior Engineer?”
Naive RAGanswers from the confidential HR file. A silent leak.
Why this matters
- HR, M&A, legal and client files in one index
- Regulators expect provable access control
- Permissions change daily
02 · Why this is hard
A SharePoint document's real audience is buried under layers that change constantly
Constant drift
People join, leave, change projects. Last month's index is wrong today.
Filtering after retrieval fails
Retrieve-then-filter leaks. Trimming must happen inside the index.
03 · The approach
Permissions are enforced twice: when documents are indexed, and when questions are asked
Index boundary
At sync time
Only documents with resolvable permissions are indexed, carrying their access list with them. Unresolvable: excluded and reported.
Query boundary
At question time
The filter is built server-side, from the verified identity of whoever asks. Unknown identity: zero results.
04 · Architecture
SharePoint stays authoritative
Microsoft 365 · authoritative source
SharePointdocuments + ACLs
Entra IDgroups & membership
Snowflake · governed knowledge
Document chunksparsed, incremental
allowed_principalsper doc: group OIDs, never expanded
entitlementsper user: transitive group closure
Curation layer
Cortex Search indexfilterable attribute: allowed_principals
auditone row per query
Consumption
Verified identitieschat app · SSO token
Snowflake agentCURRENT_USER
Permitted results only
No copies. Citations link back and a click re-enforces the live SharePoint ACL.
Permission changes propagate in ~2 minutes, with no re-indexing.
Every query audited: who asked what, and which documents answered.
05 · Under the hood
From document to grounded answer
1 · Understand
PARSE_DOCUMENT: PDF, Word, PowerPoint, to structured markdown, tables intact
2 · Chunk
Structure-aware, overlapping, title-prefixed
3 · Govern
Chunks join their access lists, fail closed
4 · Embed + index
Managed embeddings, hybrid retrieval, reranking
5 · Answer
Grounded answers with citations, from the permitted subset only
No external vector database. Everything inside Snowflake.
06 · Design guarantees
Guarantee and mechanism
GuaranteeMechanism
Fail closed everywhereNo permissions: not indexed. Unknown user: zero results.
No names in the indexObject IDs only, never names.
Server-side filter onlyClient filters rejected.
Groups stay groupsGroups referenced, never expanded: fast revocation.
Everything auditedOne audit row per query.
No governance surprisesExcluded content is reported, never silent.
Strictly additive deploymentPrefixed, monitored, one-command teardown.
Talk to your documents. Safely.
Let's look at your document estate together. bi concepts is a Snowflake Select Partner, based in Switzerland.