Solution · Retrieval-augmented generation

The model is the smallest part.

Retrieval-augmented generation that settles who may see a passage, which version wins, and what it joins to, before any model runs.

Query console
running as field.tech@customer.com
enforced in query
Unit will not hold pressure after the firmware update. What is the approved fix?
acl_principals ∩ caller groupsproduct_model = QX-4400authority_level ≥ Approvedsuperseded_by is null
Corpus
99,482
After identity
41,206
After metadata
1,873
Handed to model
5
Candidate chunkDenseSparseRecencyAuthorityFinal
QX-4400 Service Bulletin SB-118 §3.2 · approved · Mar 20260.810.740.961.000.94
Ticket #48221 · resolution · verified fix · Feb 20260.790.680.910.800.86
QX-4400 Manual r4 · superseded · 20230.880.710.100.200.31
Highest similarity, lowest rank. Authority decided, not the vector.
Why basic RAG stalls

Similarity is not judgment.

A vector index finds text that looks like the question. It does not know who is asking, what is current, or what it means in your business.

Who may see it

Basic RAG: a contractor gets a paragraph they cannot open.

DataGOL: permissions enforced inside the query.

Which version wins

Basic RAG: a 2023 manual ranks beside this quarter's revision.

DataGOL: recency and authority outrank similarity.

What it joins to

Basic RAG: quotes the contract, cannot name the account.

DataGOL: documents and warehouse answer together.

Retrieval without permissions is not search. It is a leak with good latency.
Architecture

Gate, then rank, then ground. Basic RAG only embeds.

Everything expensive happens once, at ingest. Every question runs against the surface this caller may reach.

AI-Native Data and Agents Platform | DataGOLBASIC RAGsimilarity onlyQuestionEmbedNearest neighboursAnswerNo identity. No date. No authority. No join. Every question answered in isolation.DATAGOL · INGESTruns once, per sourceSourcesACLs capturedParse and chunkper document classWrite metadataauthority, version, entitiesEmbeddense and sparseIndexplus graphONE INDEX, PERMISSIONS AND AUTHORITY ON EVERY CHUNKDATAGOL · QUERYruns per questionQuestionas this callerIdentity and metadata gaterestricted chunks never candidatesHybrid recallvector plus keywordRerankrecency and authorityGround and citeto the chunkBoth decisions are made before generation, by the index, against rules you wrote once.
01 · Permissions

The document brings its permissions with it.

Source ACLs are written onto every chunk and enforced inside the query. A restricted chunk is never a candidate.

Post-filtering
Too late

Retrieve first, filter after. The top-k is half empty and nobody can tell why.

Prompt instruction
Not enforcement

A sentence in the prompt is not a control. It cannot be audited.

Predicate in the query
Enforcement

The restricted chunk is never a candidate. Nothing to leak.

Index settings · Support KB
chunking by document class
Document classStrategyWindow
Policies and manualsSemantic~700 tok
Spec sheetsLayout awareper table
ContractsClause splitper clause
Support ticketsWhole record1 per ticket
Call transcriptsSpeaker turn~400 tok
parent expansion on15% overlapheadings prependedtables never split
02 · Chunking

Chunked the way each document is read.

One window for every document is a decision too, and usually the wrong one. Each class gets its own strategy.

03 · Recall and ranking

The highest similarity score loses.

Hybrid recall finds meaning and exact strings. Recency and authority then rank on what a vector cannot know.

Retrieval mix · Support KB
hybrid recall, then rerank
Recall
Dense, vector
0.65
Sparse, BM25
0.35
The sparse share keeps a part number or clause reference from being outvoted by a paragraph that merely sounds similar.
Rerank on
last_modifiedauthority_levelsupersedesproduct_modeldoc_class
Written onto every chunk at ingest. A field the index does not carry is a question the agent cannot be asked.
Support agent
grounded · 2 sources, 1 withheld
Re-run regulator calibration in service mode before returning the unit SB-118 §3.2. Resolved the same way on ticket #48221 Ticket 48221. Do not follow Installation Manual r4 superseded, withheld.
Cited chunk · SB-118 §3.2
The update does not rewrite the stored offset; units returned without recalibration will present as failing to hold pressure at 40 psi.
page 14 · approved by M. Rivera · 2026-03-04 · authority: Service Engineering
04 · Grounding

Every claim traces to a chunk.

Citations point to the passage, not the file. A demoted source is named, never averaged away.

05 · Graph and join

The clause and the record, in one table.

Entities resolved at ingest join a contract clause to the account, its ARR and its open tickets.

Renewals review
2 workbooks + 1 knowledge base
Which accounts renewing this quarter have an auto-renewal clause and more than five open P1 tickets?
AccountRenewalARRAuto-renewOpen P1Contract source
Acme Industrial14 Oct 2026$412,000Yes · 60 days9MSA_2024.pdf · §11.3
Northwind Logistics02 Nov 2026$286,500Yes · 90 days7Northwind_MSA.pdf · §9.1
Halcyon Foods28 Nov 2026$198,000Yes · 30 days6Halcyon_Order_Form.pdf · §4
Cobalt Systems19 Dec 2026$540,000No11excluded by clause filter
See the retrieval trace on your own documents.
Book a demo
Under the screens

Built in, not bolted on.

Ingest
Index
Retrieve
Join and act
Govern
Prove
Rollout

Four weeks to an answer you would put in front of a customer.

What takes time is not the index. It is agreeing what is authoritative.

Week 1
Content and access audit

Who may see what, and which version wins.

Week 2
Index and metadata build

Permissions and authority on every chunk.

Week 3
Pilot agent live

One team, one workflow, their tool.

Week 4
Measure and expand

Golden set scored, next corpus queued.

The test is not whether the first agent answers well. It is whether the second one was cheaper to ground.
Start with your corpus
Bring one team and the documents they argue about.

You see the retrieval trace on every answer from week one.

DataGOL · The Retrieval Layer
Solution · Retrieval-augmented generation

The model is the smallest part.

Retrieval-augmented generation that settles who may see a passage, which version wins, and what it joins to, before any model runs.

Query console
running as field.tech@customer.com
enforced in query
Unit will not hold pressure after the firmware update. What is the approved fix?
acl_principals ∩ caller groupsproduct_model = QX-4400authority_level ≥ Approvedsuperseded_by is null
Corpus
99,482
After identity
41,206
After metadata
1,873
Handed to model
5
Candidate chunkDenseSparseRecencyAuthorityFinal
QX-4400 Service Bulletin SB-118 §3.2 · approved · Mar 20260.810.740.961.000.94
Ticket #48221 · resolution · verified fix · Feb 20260.790.680.910.800.86
QX-4400 Manual r4 · superseded · 20230.880.710.100.200.31
Highest similarity, lowest rank. Authority decided, not the vector.
Why basic RAG stalls

Similarity is not judgment.

A vector index finds text that looks like the question. It does not know who is asking, what is current, or what it means in your business.

Who may see it

Basic RAG: a contractor gets a paragraph they cannot open.

DataGOL: permissions enforced inside the query.

Which version wins

Basic RAG: a 2023 manual ranks beside this quarter's revision.

DataGOL: recency and authority outrank similarity.

What it joins to

Basic RAG: quotes the contract, cannot name the account.

DataGOL: documents and warehouse answer together.

Retrieval without permissions is not search. It is a leak with good latency.
Architecture

Gate, then rank, then ground. Basic RAG only embeds.

Everything expensive happens once, at ingest. Every question runs against the surface this caller may reach.

AI-Native Data and Agents Platform | DataGOLBASIC RAGsimilarity onlyQuestionEmbedNearest neighboursAnswerNo identity. No date. No authority. No join. Every question answered in isolation.DATAGOL · INGESTruns once, per sourceSourcesACLs capturedParse and chunkper document classWrite metadataauthority, version, entitiesEmbeddense and sparseIndexplus graphONE INDEX, PERMISSIONS AND AUTHORITY ON EVERY CHUNKDATAGOL · QUERYruns per questionQuestionas this callerIdentity and metadata gaterestricted chunks never candidatesHybrid recallvector plus keywordRerankrecency and authorityGround and citeto the chunkBoth decisions are made before generation, by the index, against rules you wrote once.
01 · Permissions

The document brings its permissions with it.

Source ACLs are written onto every chunk and enforced inside the query. A restricted chunk is never a candidate.

Post-filtering
Too late

Retrieve first, filter after. The top-k is half empty and nobody can tell why.

Prompt instruction
Not enforcement

A sentence in the prompt is not a control. It cannot be audited.

Predicate in the query
Enforcement

The restricted chunk is never a candidate. Nothing to leak.

Index settings · Support KB
chunking by document class
Document classStrategyWindow
Policies and manualsSemantic~700 tok
Spec sheetsLayout awareper table
ContractsClause splitper clause
Support ticketsWhole record1 per ticket
Call transcriptsSpeaker turn~400 tok
parent expansion on15% overlapheadings prependedtables never split
02 · Chunking

Chunked the way each document is read.

One window for every document is a decision too, and usually the wrong one. Each class gets its own strategy.

03 · Recall and ranking

The highest similarity score loses.

Hybrid recall finds meaning and exact strings. Recency and authority then rank on what a vector cannot know.

Retrieval mix · Support KB
hybrid recall, then rerank
Recall
Dense, vector
0.65
Sparse, BM25
0.35
The sparse share keeps a part number or clause reference from being outvoted by a paragraph that merely sounds similar.
Rerank on
last_modifiedauthority_levelsupersedesproduct_modeldoc_class
Written onto every chunk at ingest. A field the index does not carry is a question the agent cannot be asked.
Support agent
grounded · 2 sources, 1 withheld
Re-run regulator calibration in service mode before returning the unit SB-118 §3.2. Resolved the same way on ticket #48221 Ticket 48221. Do not follow Installation Manual r4 superseded, withheld.
Cited chunk · SB-118 §3.2
The update does not rewrite the stored offset; units returned without recalibration will present as failing to hold pressure at 40 psi.
page 14 · approved by M. Rivera · 2026-03-04 · authority: Service Engineering
04 · Grounding

Every claim traces to a chunk.

Citations point to the passage, not the file. A demoted source is named, never averaged away.

05 · Graph and join

The clause and the record, in one table.

Entities resolved at ingest join a contract clause to the account, its ARR and its open tickets.

Renewals review
2 workbooks + 1 knowledge base
Which accounts renewing this quarter have an auto-renewal clause and more than five open P1 tickets?
AccountRenewalARRAuto-renewOpen P1Contract source
Acme Industrial14 Oct 2026$412,000Yes · 60 days9MSA_2024.pdf · §11.3
Northwind Logistics02 Nov 2026$286,500Yes · 90 days7Northwind_MSA.pdf · §9.1
Halcyon Foods28 Nov 2026$198,000Yes · 30 days6Halcyon_Order_Form.pdf · §4
Cobalt Systems19 Dec 2026$540,000No11excluded by clause filter
See the retrieval trace on your own documents.
Book a demo
Under the screens

Built in, not bolted on.

Ingest
Index
Retrieve
Join and act
Govern
Prove
Rollout

Four weeks to an answer you would put in front of a customer.

What takes time is not the index. It is agreeing what is authoritative.

Week 1
Content and access audit

Who may see what, and which version wins.

Week 2
Index and metadata build

Permissions and authority on every chunk.

Week 3
Pilot agent live

One team, one workflow, their tool.

Week 4
Measure and expand

Golden set scored, next corpus queued.

The test is not whether the first agent answers well. It is whether the second one was cheaper to ground.
Start with your corpus
Bring one team and the documents they argue about.

You see the retrieval trace on every answer from week one.

DataGOL
Solution · Retrieval-augmented generation

The model is the smallest part.

Retrieval-augmented generation that settles who may see a passage, which version wins, and what it joins to, before any model runs.

Query console
running as field.tech@customer.com
enforced in query
Unit will not hold pressure after the firmware update. What is the approved fix?
acl_principals ∩ caller groupsproduct_model = QX-4400authority_level ≥ Approvedsuperseded_by is null
Corpus
99,482
After identity
41,206
After metadata
1,873
Handed to model
5
Candidate chunkDenseSparseRecencyAuthorityFinal
QX-4400 Service Bulletin SB-118 §3.2 · approved · Mar 20260.810.740.961.000.94
Ticket #48221 · resolution · verified fix · Feb 20260.790.680.910.800.86
QX-4400 Manual r4 · superseded · 20230.880.710.100.200.31
Highest similarity, lowest rank. Authority decided, not the vector.
Why basic RAG stalls

Similarity is not judgment.

A vector index finds text that looks like the question. It does not know who is asking, what is current, or what it means in your business.

Who may see it

Basic RAG: a contractor gets a paragraph they cannot open.

DataGOL: permissions enforced inside the query.

Which version wins

Basic RAG: a 2023 manual ranks beside this quarter's revision.

DataGOL: recency and authority outrank similarity.

What it joins to

Basic RAG: quotes the contract, cannot name the account.

DataGOL: documents and warehouse answer together.

Retrieval without permissions is not search. It is a leak with good latency.
Architecture

Gate, then rank, then ground. Basic RAG only embeds.

Everything expensive happens once, at ingest. Every question runs against the surface this caller may reach.

AI-Native Data and Agents Platform | DataGOLBASIC RAGsimilarity onlyQuestionEmbedNearest neighboursAnswerNo identity. No date. No authority. No join. Every question answered in isolation.DATAGOL · INGESTruns once, per sourceSourcesACLs capturedParse and chunkper document classWrite metadataauthority, version, entitiesEmbeddense and sparseIndexplus graphONE INDEX, PERMISSIONS AND AUTHORITY ON EVERY CHUNKDATAGOL · QUERYruns per questionQuestionas this callerIdentity and metadata gaterestricted chunks never candidatesHybrid recallvector plus keywordRerankrecency and authorityGround and citeto the chunkBoth decisions are made before generation, by the index, against rules you wrote once.
01 · Permissions

The document brings its permissions with it.

Source ACLs are written onto every chunk and enforced inside the query. A restricted chunk is never a candidate.

Post-filtering
Too late

Retrieve first, filter after. The top-k is half empty and nobody can tell why.

Prompt instruction
Not enforcement

A sentence in the prompt is not a control. It cannot be audited.

Predicate in the query
Enforcement

The restricted chunk is never a candidate. Nothing to leak.

Index settings · Support KB
chunking by document class
Document classStrategyWindow
Policies and manualsSemantic~700 tok
Spec sheetsLayout awareper table
ContractsClause splitper clause
Support ticketsWhole record1 per ticket
Call transcriptsSpeaker turn~400 tok
parent expansion on15% overlapheadings prependedtables never split
02 · Chunking

Chunked the way each document is read.

One window for every document is a decision too, and usually the wrong one. Each class gets its own strategy.

03 · Recall and ranking

The highest similarity score loses.

Hybrid recall finds meaning and exact strings. Recency and authority then rank on what a vector cannot know.

Retrieval mix · Support KB
hybrid recall, then rerank
Recall
Dense, vector
0.65
Sparse, BM25
0.35
The sparse share keeps a part number or clause reference from being outvoted by a paragraph that merely sounds similar.
Rerank on
last_modifiedauthority_levelsupersedesproduct_modeldoc_class
Written onto every chunk at ingest. A field the index does not carry is a question the agent cannot be asked.
Support agent
grounded · 2 sources, 1 withheld
Re-run regulator calibration in service mode before returning the unit SB-118 §3.2. Resolved the same way on ticket #48221 Ticket 48221. Do not follow Installation Manual r4 superseded, withheld.
Cited chunk · SB-118 §3.2
The update does not rewrite the stored offset; units returned without recalibration will present as failing to hold pressure at 40 psi.
page 14 · approved by M. Rivera · 2026-03-04 · authority: Service Engineering
04 · Grounding

Every claim traces to a chunk.

Citations point to the passage, not the file. A demoted source is named, never averaged away.

05 · Graph and join

The clause and the record, in one table.

Entities resolved at ingest join a contract clause to the account, its ARR and its open tickets.

Renewals review
2 workbooks + 1 knowledge base
Which accounts renewing this quarter have an auto-renewal clause and more than five open P1 tickets?
AccountRenewalARRAuto-renewOpen P1Contract source
Acme Industrial14 Oct 2026$412,000Yes · 60 days9MSA_2024.pdf · §11.3
Northwind Logistics02 Nov 2026$286,500Yes · 90 days7Northwind_MSA.pdf · §9.1
Halcyon Foods28 Nov 2026$198,000Yes · 30 days6Halcyon_Order_Form.pdf · §4
Cobalt Systems19 Dec 2026$540,000No11excluded by clause filter
See the retrieval trace on your own documents.
Book a demo
Under the screens

Built in, not bolted on.

Ingest
Index
Retrieve
Join and act
Govern
Prove
Rollout

Four weeks to an answer you would put in front of a customer.

What takes time is not the index. It is agreeing what is authoritative.

Week 1
Content and access audit

Who may see what, and which version wins.

Week 2
Index and metadata build

Permissions and authority on every chunk.

Week 3
Pilot agent live

One team, one workflow, their tool.

Week 4
Measure and expand

Golden set scored, next corpus queued.

The test is not whether the first agent answers well. It is whether the second one was cheaper to ground.
Start with your corpus
Bring one team and the documents they argue about.

You see the retrieval trace on every answer from week one.

DataGOL