Spaces:
Running
Running
Commit History
-
feat(k8s): deploy api to kind multinode cluster and verify with adversarial red-teaming simulator
Saibalaji Namburi5553f71
-
feat(observability): upgrade Langfuse tracing per langfuse/skills best practices
Saibalaji NamburiApplied all baseline requirements from github.com/langfuse/skills: - Descriptive trace name: 'ticket-triage' (was generic 'triage') - Explicit trace input: only ticket_text + channel (not all function args) - PII masking: email/phone/SSN/IBAN/card stripped before sending to Langfuse - Singleton client: one Langfuse() per process (not one per request) - flush via client.flush() not trace.flush() (SDK v3 best practice) - agent_span accepts input_data param for meaningful span inputs - Generation names use hyphens: 'gemma3:4b-generation' (was underscore) - Prompt messages and completions PII-masked before logging - Tags use bounded cardinality: [tier, channel] not tenant_id5e56072 -
feat(phase11): Langfuse LLM tracing + Constitutional Policy Engine
Saibalaji NamburiLangfuse Observability (src/monitoring/langfuse_tracer.py): - TriageTrace: trace/span/generation hierarchy per triage request - Graceful no-op degradation when LANGFUSE_PUBLIC_KEY not set - LiteLLM callback registration (auto-traces all LLM calls) - Quality scores: constitutional_compliance, resolution_quality, rag_grounding - Wired into API lifespan startup and triage background task Constitutional Policy Engine (src/responsible_ai/constitutional_policy.py): - 7 constitutional rules: PII_PROTECTION, AI_IDENTITY, TOXICITY_GUARD, NO_COMMITMENTS, SCOPE_LIMITATION, LANGUAGE_CONSISTENCY, COMPETITOR_NEUTRAL - 3 severity levels: CRITICAL (block), VIOLATION (regenerate/redact), WARNING (warn) - Fast regex-based checks (<5ms) for all rules - PolicyVerdict with score (0-1), action, violations, warnings, safe_fallback - Critical violations auto-trigger HITL escalation - Runtime rule addition/removal support Tests: 38 new tests, all passingbf95443