System Status & Configuration
Platform runtime configuration, API provider keys, ADR architecture standards & resource boundaries
Infrastructure Architecture
Core platform storage engines, distributed queue brokers & caching backends
Database Engine (Cloud Pooler)
POSTGRES_URL — PostgreSQL (asyncpg) on Supabase Cloud Pooler (aws-0-ap-southeast-1)
Redis Cache & Distributed Broker
REDIS_URL — L1 memory + L2 Redis cache & Celery async queue execution
Vector Database & Semantic Index
QDRANT_URL / SQLite-VSS — MinHash LSH shingling & dense embedding deduplication
Standalone Local Container Execution Mode
Multiprocess worker pool with zero external token dependency & deterministic test harness
Source-Grounded Engine Standards
Enforcing authentic retrieval-grounded generation over hallucinated pure synthetic data
Default Generation Mode
GENERATION_MODE=strict_source — Pure synthetic generation is strictly disabled unless explicitly opted in
strict_sourceSeedless Generation Guard
ALLOW_SEEDLESS_GENERATION=false — Generation without grounded domain reference seeds is rejected
Disabled (Protected)Topological Coverage Planner
COVERAGE_PLANNER_ENABLED=true — Plans domain taxonomy distribution prior to LLM extraction calls
EnabledValidation Strictness
VALIDATION_STRICTNESS=standard — Syntactic, schema, and semantic sanity checks enforced on all batches
standardWhy does this guard exist? Pure synthetic text generated without reference seed documents frequently invents fictitious legal sections, fabricated statutory penalties, and hallucinated medical dosages.
When enabled (ALLOW_SEEDLESS_GENERATION=false), the pipeline strictly requires an authoritative reference corpus (e.g., CGST Act PDFs, medical case studies, or validated codebases) before initiating LLM teacher-forcing. If seed retrieval fails or yields zero matches, generation is automatically halted rather than allowing the model to hallucinate unverified statutory claims from vacuum.