Engineered from ground zero for mission-critical document research, verifiable citations, and mathematically grounded synthesis across massive heterogeneous document vaults.
Standard RAG relies on naive vector similarity which frequently hallucinates or misses exact part numbers, contract clauses, and proper nouns. KURIPP executes Reciprocal Rank Fusion directly inside PostgreSQL, merging high-dimensional HNSW cosine vectors with BM25 trigram full-text weights in a single mathematical pass.
| Architecture Dimension | Generic Naive RAG | KURIPP Geological System |
|---|---|---|
| Retrieval Mechanism | Cosine Distance Only (Prone to semantic drift) | Hybrid RRF: 1536D HNSW + pg_trgm (Single SQL pass) |
| Browser Security | JWT in localStorage (Vulnerable to XSS extraction) | Strict HttpOnly SameSite=Lax Cookie + In-Memory Token |
| Model Vendor Lock-in | Hardcoded OpenAI / Anthropic SDK | OpenRouter Dynamic Resolvers (Llama 3.3, Gemini 2.0, Qwen) |
| Contract Diffing | Raw character git diffs (No semantic understanding) | Semantic Clause Classification & Liability Risk Delta |
| Evaluation Rigor | Subjective manual spot checks | Automated Continuous Benchmark (92% Groundedness Target) |
| Ingestion Pipeline | Blocking client-side script | Upstash Redis 7 + BullMQ Async Worker Queues |
Create your primary workspace, index your document archives, and run your first hybrid search in minutes.