ragi-instant

🇮🇩 Bahasa Indonesia

Production-grade RAG for regulatory compliance. Hybrid search, reranking, routing, hallucination detection — all in one pipeline.

---

Ragi-Instant is a RAG (Retrieval-Augmented Generation) system built for regulatory and compliance intelligence. It goes beyond the tutorial-level embed → search → generate loop with a production-grade pipeline: query rewriting, intent routing, hybrid retrieval, cross-encoder reranking, context compression, citation-grounded structured output, and hallucination checking — all traced and evaluated automatically.

Pipeline

User Query
    │
    ▼
Query Rewriting (LLM)     "POJK terbaru pinjol" → "Peraturan OJK peer-to-peer lending 2024 2025"
    │
    ▼
Intent Router (LLM)       regulation_lookup | definition | comparison | obligation_check
    │
    ▼
Hybrid Search             Dense (pgvector cosine) + Sparse (PostgreSQL FTS BM25) → RRF fusion
    │
    ▼
Cohere Rerank v3          Top-40 → top-5 final results
    │
    ▼
[Context Compression]     LLM summarization preserving legal facts (optional)
    │
    ▼
LLM Generation            GPT-4o mini / Claude Haiku + Instructor structured output
    │
    ▼
Hallucination Check       Verify every claim against source context
    │
    ▼
Answer + Citations + Confidence + Related Regulations
    │
    ▼
LangFuse Trace + RAGAS Eval (async)

Key Capabilities

Query Rewriting LLM-based query expansion. "POJK terbaru pinjol" becomes "Peraturan OJK terbaru mengenai peer-to-peer lending 2024 2025" before retrieval. +15-20% retrieval quality for ambiguous queries.

Intent Routing Classifies queries into regulation_lookup, definition, comparison, or obligation_check. Each intent maps to a different search strategy (dense-only, sparse-only, or hybrid).

Hybrid Search Dense vector search (pgvector HNSW) combined with sparse keyword retrieval (PostgreSQL FTS BM25 with GIN index), fused via Reciprocal Rank Fusion. Catches both semantic matches and exact legal term hits.

Reranking Cohere Rerank v3 cross-encoder on retrieved candidates. Improves recall@5 significantly over raw vector similarity.

Context Compression LLM-based summarization that preserves numbers, percentages, article numbers, and legal definitions while removing transitions and repetition. 40-60% token reduction.

Structured Output & Citations Every answer comes with source citations pointing to the exact chunk, document, page, and section. No black-box answers — every claim is verifiable.

Hallucination Detection Post-generation verification step. Each claim in the answer is checked against the source context. Returns hallucination score and flags unsupported claims.

Regulatory Change Detection Upload two versions of a regulation and get a structured diff — which articles changed, what was added, what was removed — grouped by impact (HIGH/MEDIUM/LOW).

Automated Evaluation Built-in RAGAS evaluation pipeline with 30 curated Q&A pairs. Measures faithfulness, answer relevancy, and context precision. You ship with numbers, not faith.

Observability LangFuse end-to-end tracing of every step: query rewriting, routing, hybrid search, reranking, compression, generation, and hallucination check. Latency and confidence per step.

Demo

Here is a visual preview of the primary dashboard interfaces:

Startup Splash Screen	Main System Dashboard

AI Query Workspace & Citations	Automated RAGAS Benchmarks

👉 Explore the full 16-screenshot gallery with detailed pipeline step descriptions in the Ragi Instant Demo Gallery.

Installation

cd backend
pip install -e .

Quick Start

# 1. Copy env template and fill API keys
cp .env.example .env
# Required: OPENAI_API_KEY, COHERE_API_KEY

# 2. Deploy to VPS
make deploy

# 3. Upload a regulation document
curl -F "file=@POJK_10_2022.pdf" \
     -F "title=POJK No. 10 Tahun 2022" \
     http://vps:8000/api/ingest/documents

# 4. Ask a question
curl -X POST http://vps:8000/api/query \
  -H "Content-Type: application/json" \
  -d '{"question": "Apa batas maksimum bunga pinjaman online menurut OJK?"}'

# Response:
# {
#   "answer": "Berdasarkan POJK No. 10/PT. LKM/2022, batas maksimum...",
#   "citations": [{"document_title": "POJK No. 10/2022", "page": 12, "quote": "..."}],
#   "confidence": 0.89,
#   "related_regulations": ["POJK No. 22/2023"]
# }

API Reference

Endpoint	Method	Description
/health	GET	Health check
/api/ingest/documents	POST	Upload PDF/DOCX for indexing
/api/ingest/documents	GET	List all documents
/api/ingest/documents/{id}	GET	Document detail
/api/ingest/documents/{id}/chunks	GET	View indexed chunks
/api/query	POST	Full RAG query pipeline
/api/analysis/compare	POST	Compare two document versions

Benchmark

Evaluated on 30 Q&A pairs from Indonesian financial regulatory documents (POJK, PBI).

Metric	Score	Target
Faithfulness	0.89	> 0.85
Answer Relevancy	0.83	> 0.80
Context Precision	0.76	> 0.75
Avg Latency	1.27s	< 2.0s
Avg Cost/Query	<$0.01	< $0.01

Scores evaluated using GPT-4o as the judge over the curated POJK & PBI benchmark dataset.

🔍 Production-Grade Observability

In legal and regulatory compliance, deploying a RAG pipeline blindly and relying on "hallucination luck" is a major liability. Ragi Instant takes an engineering-first approach: we measure and trace every single token, latency, cost, and intermediate step.

Using Langfuse, the entire hybrid pipeline is transparently traced:

End-to-End Tracing (Spans)	Context Compression & Costs

We trace and audit every single modular step:

• Query Rewriting: Visualizing how ambiguous user queries expand.
• Query Intent Routing: Inspecting LLM intent classification confidence and routing targets.
• Hybrid Search: Dense pgvector matches & sparse PostgreSQL FTS BM25 database logs.
• Cohere Reranking: Evaluating cross-encoder scores and candidate ordering.
• Context Compression: Auditing token savings and LLM summarization context before generation.
• Hallucination Check: Factual checks against source context to identify unsupported claims.

👉 View full detailed traces for each step in the Langfuse Observability & Trace Gallery.

Stack

Layer	Technology
Document Parsing	Docling (layout-aware PDF + table extraction)
Chunking	LlamaIndex SemanticSplitterNodeParser
Embedding	OpenAI text-embedding-3-small
Vector Store	pgvector (HNSW index)
Keyword Search	PostgreSQL FTS (BM25, GIN index)
Backend API	FastAPI (async)
Frontend	Next.js 15 + shadcn/ui + Tailwind CSS
Tracing	LangFuse
Evaluation	RAGAS

License

MIT © Muhammad Faisal Affan