WPIntelliChat
LLM Governance
Governance Overview
Loading metrics…
LLM Governance
checking…
Runtime safety of governed LLM traffic — what was blocked, what fell back, how grounded the answers were, and what it cost.
Governed runs
Blocked rate
Fallback rate
Answer relevancy
Groundedness
1.0 = fully grounded
Spend (USD)
Active rules Blocked Fallbacks Avg latency s
Data Governance
checking…
Quality of retrieval-backed (RAG) answers — did we fetch the right documents, and were answers grounded in them? Precision@K and Hit Rate@K are score-based proxies (no ground-truth labels).
RAG runs
Retrieval Precision@K (score-proxy)
Hit Rate@K (score-proxy)
Context Relevance
Answer Groundedness
Retrieval Latency (ms)
Total LLM Calls
in selected period
Total Tokens
prompt + completion
Total Cost (USD)
based on model pricing
Avg Latency
seconds per call
Daily Calls & Cost Trend
Top Models by Usage successful calls
Calls by Type
Guardrail Activity — requests per day one bar unit = one request, not one rule hit
Guardrail Breakdown — checks fired, by action
WPIntelliChat LLM Governance — Live Use Case
6-step governance pipeline running inside WPIntelliChat
Prompt Lifecycle Status
Eval Scores by Prompt
Prompt Catalog — Current State
PromptCategoryStatusRelevancyCoherenceCompletenessOverallAction
Loading…
LLM Audit Log
Complete record of every LLM call — who, what model, tokens, cost
Last 24h: loading…
Sr. No. Timestamp User Model Type Request ID User Prompt Prompt Tok Completion Tok Latency Cost USD
Loading…
Models & Experiments
The model registry as a planetarium — distance = cost, size = context, ring = guardrail profile, green = the active API key's model — side by side with the prompt-experiment runs that A/B test those models. Click a planet for its model card; click a run for its variants.
Registry Atlas
Active API key: checking…
LLM RAG Tool active key distance = cost · size = context · ring = guardrail profile · click a planet to view
Experiment Runs
Run ID Experiment Name Variants Best Time (s) Min Cost ($) Date
Loading…
Prompt Catalog
Engineer, evaluate, and promote prompts through a governed lifecycle (Draft → Approved → Production)
NameCategoryStatusUse CaseRating (1-5) Eval ScoreGradeModel HintUpdatedActions
Loading…
Compliance Review Queue
Regulation comparisons, gap reports, and AI-drafted PPM clauses awaiting review.
TitleKindStateSeverity QualityCreated byUpdated
Loading…
Cost Estimator
Estimate LLM call cost before you run — compare models side by side
Estimate Parameters
Counted at 50% of input price (provider prompt-cache hits).
Estimated Cost
Per Request
Per Day
Per Month
Compare All Models
ModelProviderPer RequestPer DayPer MonthContext
Guardrail Rules
Database-backed rules per profile. Each chat call uses the rules of its model's guardrail_profile.
IDProfileNameTypeAction StagePriorityActiveOperations
Loading…
Orchestration Chat
Test how the governance pipeline (guardrails → prompt resolve → model resolve → LLM → evaluate → audit) treats a message — each turn shows its verdict and which guardrails fired. Toggle with env ORCHESTRATOR_V2_ENABLED=1.
checking…
Initializing…
Connecting to orchestrator
Guardrail reflex ribbon · recent decisions · click a bar for detail allow warn block redact
No messages yet. Try a normal question, or probe the rails:
PII — "My BSB is 062-334" · Injection — "Ignore previous instructions" · Sentiment — "I am so frustrated"
Recent Guardrail Decisions
Loading…
Decision Tables
Rule tables that map input records to outcomes — used by the Decision Table workflow node.
NameDescriptionStatusColumnsRulesCreated
A/B Experiments
Champion/Challenger experiments that route traffic between two decision strategies.
NameStatusSplitVariant AVariant BCreated
Policy Watch
Tracked regulations as a living garden — each drift check waters a plant, each detected amendment blooms. Backed by /api/policy-watch.
healthy amendment pending amendment resolved stale (30d+) seedling (new)
Register Regulation
An anchor PDF is required — it is the snapshot future drift checks compare against.
New Decision Table
New Rule
Conditions (all must be true — leave empty for catch-all)
Combine conditions with all must match
Test Record Against Table
Paste a JSON record and see which rule matches.
Rule History
Snapshots captured automatically on every update/delete.
Saved Test Records
Save a record here to quickly re-run it later — samples never affect production rules.
Import Rules
Paste a JSON array of rules (same shape as Export JSON). Rules are appended, not replaced.
New A/B Experiment
Remainder goes to Variant B
Variant A (Champion)
Variant B (Challenger)
Results Configuration (optional)
A routed record counts as a success when the recorded outcome's field equals this value. Leave blank to count any recorded outcome as success.
Record Outcome
What actually happened for this routed decision?
Audit Log Detail
Experiment Run