Attribution
Which passage supported which claim, at which version of the document.
Reference architecture
Agentic RAG Framework
A permission-aware architecture for enterprise knowledge ingestion, retrieval, grounding, orchestration, access control, evaluation, and observability.
A production-grade RAG pattern teams can adopt, extend and operate without us.
Architecture evidence
What the architecture defines
- RAG reference architecture
- Permission model
- Evaluation loop design
CLIENT CONTEXT
Enterprises need knowledge systems that answer accurately, respect permissions, and can be observed and improved in production.
BUSINESS PROBLEM
Naive RAG implementations leak data, hallucinate, and degrade silently.
CONSTRAINTS
- Enterprise permissions that retrieval has to respect, not re-implement
- Knowledge fragmented across systems and formats
- Degradation that naive RAG does not surface until someone complains
ARCHITECTURE & APPROACH
Engineered a reference architecture covering ingestion, indexing, permission-aware retrieval, grounding, agent orchestration, evaluation, and observability.
WHAT CAPTIVOLT DELIVERED
Reference architecture · permission-aware retrieval model · grounding and traceability design · evaluation and observability loop.
ENGINEERING DECISIONS
- Permission-aware retrieval rather than filtering results afterwards
- Grounding and citation treated as requirements, not features
- Evaluation and observability wired in from the first build
EVIDENCE
Architecture walkthrough available.
In detail
One question, traced through the framework.
01 · Architecture trace
The complete trace
Seven stages, described mechanically: what each does to the data, what it emits, and what goes wrong when it is missing. A worked business question runs through the same stages on the Agentic RAG Accelerator page; this is the architecture underneath it.
STAGE 01
Question
WHAT HAPPENS
Classified and routed before anything is searched: an identifier goes to exact lookup, a conceptual question to semantic search, most real questions to both.
OUTPUT
Emits the routing decision and the caller’s identity.
IF MISSING
Skipped, the system searches a vector index for a contract number and returns things that look like one.
Select a stage above for focused reading, or expand the complete technical reference.
01QuestionRoute + identity
WHAT HAPPENS
Classified and routed before anything is searched: an identifier goes to exact lookup, a conceptual question to semantic search, most real questions to both.
OUTPUT
Emits the routing decision and the caller’s identity.
IF MISSING
Skipped, the system searches a vector index for a contract number and returns things that look like one.
02RetrievalPermissioned candidates
WHAT HAPPENS
Hybrid (keyword and vector over the same corpus), executed under the caller’s permissions rather than a service account.
OUTPUT
Emits the candidate set with each document’s source, version and classification.
IF MISSING
Skipped as a permission step, the system is one well-phrased question from an access incident.
03RerankingOrdered shortlist
WHAT HAPPENS
A slower, more accurate pass re-scores the shortlist against the actual question.
OUTPUT
Emits the ordering and the scores that produced it.
IF MISSING
Skipped, the top of the list is whatever first-pass similarity happened to like, which is the part most worth fixing.
04ContextGrounding package
WHAT HAPPENS
The passages that earn their place, inside a budget, with metadata intact: effective date, supersession, classification.
OUTPUT
Emits the assembled context and what was dropped for space.
IF MISSING
Skipped, more context is sent and the relevant passage competes with nine irrelevant ones.
05AnswerSupported response
WHAT HAPPENS
Generated from the supplied passages, with declining to answer treated as a correct outcome when they do not support one.
OUTPUT
Emits the answer and the passages it drew on.
IF MISSING
Skipped as a constraint, the model fills the gap fluently.
06CitationClaim attribution
WHAT HAPPENS
Each claim bound to its passage at generation time rather than matched back afterwards.
OUTPUT
Emits claim-level attribution with document, section and version.
IF MISSING
Skipped, you get a plausible answer and a bibliography, which is not the same thing.
07EvaluationScores + learning
WHAT HAPPENS
Groundedness and retrieval precision scored on the run, with failures and user corrections routed into the evaluation set.
OUTPUT
Emits the scores and what entered the suite.
IF MISSING
Skipped, retrieval quality decays silently as documents are added, superseded and re-permissioned underneath it.
02 · Governed answer record
What the trace is retained for
The trace is not a debugging convenience. It is the artefact that makes an answer usable in a decision somebody has to defend.
GOVERNED ANSWER RECORD
Retained trace
Caller · routing · corpus state · candidates · context · claims · citations · evaluation
Entitlement
Whose permissions the retrieval ran under, and what was excluded because of them.
Reproducibility
The corpus version and the routing decision, so the same question can be re-asked against the same state.
Improvement
What the evaluation scored and what a user corrected: the input to the next dataset rather than to a backlog.
Defensible enterprise useEvidence for decisions · access review · replay · controlled improvement
OUTCOME
A production-grade pattern that teams can adopt, extend, and operate independently.
WHAT THE CLIENT OWNS NOW
- The reference architecture
- The permission-aware retrieval model
- The grounding and traceability design
- The evaluation and observability loop
RELATED SOLUTION
Explore the capabilities behind the architecture.
Agentic RAG Accelerator · VeriCore
Want to see the artefacts?
Start with the work most relevant to your initiative.
Anonymised artefacts and reference discussions are available under NDA where client permission allows.