Worked sample · not a client case study
What a Context Integrity Assessment actually produces.
This fictional policy-support RAG example shows the structure, evidence and decision boundaries of the paid assessment. It does not claim a real customer result.
- workflow
- policy-support RAG assistant
- risk
- stale policy retrieval
- environment
- staging
- service
- Context Integrity Assessment
- output
- fit map + evidence + implementation estimate
Executive finding
The failure is upstream of generation.
The assistant can produce fluent answers from policy pages that are semantically relevant but operationally stale. The primary insertion point is between retrieval and answer generation, where candidate context can be evaluated before it becomes trusted prompt context.
Workflow map
Recommended insertion point.
The assessment makes the integration seam explicit rather than simply recommending “use FreshContext.”
| Stage | Current behavior | Assessment treatment |
|---|---|---|
| Policy corpus | Pages carry mixed update dates and source authority | Preserve source/date metadata where available |
| Retriever | Ranks primarily on semantic relevance | Keep retriever unchanged |
| Context gate | No explicit integrity decision | Insert FreshContext here |
| Generator | Receives top retrieved passages as trusted context | Receive only context plus explicit decisions/reasons |
| Answer | May cite stale policy as current | Application decides how to handle verify/refresh/background outcomes |
Risk register
Observed context-integrity risks.
A real assessment ties each concern to evidence, operational effect and residual risk.
| Risk | Evidence | Operational effect | FreshContext treatment |
|---|---|---|---|
| Stale policy page | Relevant page last updated 14 months ago | Outdated guidance may dominate semantic rank | Apply source-aware temporal pressure and surface refresh/verification need |
| Unknown date | One retrieved document has no reliable publication/update date | System silently treats recency as unknown but still usable | Preserve date uncertainty instead of converting it to current |
| Weak provenance | Copied policy summary lacks canonical source | Agent may cite a derivative summary as primary | Reduce readiness and expose provenance weakness |
| Conflicting sources | Two pages describe different thresholds | Generator may synthesize contradiction | Return explicit decision/reason; caller still owns resolution |
Representative evidence
Sample decision set.
These are illustrative examples of the type of evidence supplied. They are not claims about a real customer's data.
Current canonical policy
Expected: usable as primary/supporting context where the configured source/date confidence is strong and the content is relevant.
Old but highly relevant policy
Expected: semantic relevance remains visible, but temporal decay prevents relevance alone from silently making it current.
Undated summary
Expected: date uncertainty is surfaced and can drive verify/background handling rather than an invented freshness claim.
Unavailable source
Expected: unknown/unavailable state remains unknown; it is not rewritten as healthy, current or verified.
Implementation recommendation
Move to a single-workflow pilot.
The assessment closes with a concrete commercial next step, not an open-ended consulting recommendation.
Recommended scope
One staging RAG workflow; one insertion point after retrieval; agreed metadata contract; five deterministic acceptance scenarios; operator handover.
Dependencies
Representative retrieved payloads, source/date metadata where available, staging access, technical owner, and agreement on how verify/refresh/background decisions affect generation.
- service
- Single-Workflow Integration
- typical range
- US$2,500–5,000
- acceptance
- five agreed scenarios + handover
- not included
- SLA, SSO, compliance certification, truth verification
Commercial proof
You can see the output before buying the work.
The paid assessment applies this structure to your actual workflow, evidence and constraints.