Travel workflows and seller-intelligence scenarios test the same pipeline against different contexts.
Grounded retrieval / guardrails / release logic
RAG Pipeline with Guardrails
The vector baseline for the enterprise knowledge layer: grounded answers, safety checks, benchmark traces, and release gates connected in one production-minded system. A separate implemented GraphRAG layer now tests when relationship traversal earns its extra complexity.
- Domains
- 2 corpora
- Eval cases
- 24
- Verdict
- PASS
- Guardrails
- 4 classes
- Decision
- Connect citations, confidence, input/output guardrails, and evaluation traces in one pipeline.
- Why
- Retrieval accuracy alone cannot support a release decision or explain a failure.
- Result
- A 24-case, dual-domain baseline with four guardrail classes and an explicit PASS verdict.
Retrieval quality alone does not make a feature shippable.
A grounded answer can still be risky, misleading, or hard to defend. The useful product question is whether the system can produce grounded answers, flag unsafe behavior, and explain why a release should move forward or stop.
Prompt injection, PII, toxicity, and groundedness are checked around the answer flow.
Citations, confidence, and verdict reasons make the answer defensible.
Policy thresholds turn raw eval output into a launch-review decision.
The important part is how the pieces fit together.
Domain content
Travel and seller workflow knowledge bases.
Grounding
Relevant context is selected before synthesis.
Risk checks
Inputs and outputs are checked before release.
Decision
Reports and gates produce PASS, WARN, or BLOCK.
Questions resolve to ranked committed source text.
Run npm test && npm run eval && npm run gate. The public workbench uses the source repository’s tokenizer, term-weighted ranking, injection rule set, extractive generator, groundedness calculation, and abstention boundary. It makes no live-model claim.