PRODUCTION
RAG SYSTEM

A production-grade Retrieval-Augmented Generation pipeline with hybrid BM25 + pgvector retrieval, RAGAS evaluation, LangSmith observability, and full GCP deployment via Terraform + GitHub Actions.

Launch Demo View Source Swagger UI
3×
Retrieval Modes
4
RAGAS Metrics
1024d
Embedding Dims
∞
Doc Capacity
02 — System Design

ARCHITECTURE

Ingestion, retrieval, and evaluation pipelines running on GCP Cloud Run backed by PostgreSQL + pgvector and Redis caching.

CLIENT INTERNET HTTP CORE ENGINE FastAPI Document Loader PDF · TXT → Chunks Embedder BAAI/bge-large (1024d) RRF Fusion Reciprocal Rank Fusion RAGAS Evaluator Faithfulness · Relevancy dense sparse cache POSTGRES 16 pgvector cosine similarity Cloud SQL IN-PROCESS BM25 Index rank-bm25 · TF-IDF CACHING Redis 7 OBSERVABILITY LangSmith trace · eval · monitor CI/CD GitHub Actions test → build → push → deploy Terraform · Cloud Run GCP Cloud Run serverless · auto-scale deploy
📥
Ingestion Pipeline
Upload PDF/TXT → extract → chunk (fixed or semantic) → embed with BAAI/bge-large-en-v1.5 → store in pgvector + update BM25 index → invalidate Redis cache.
🔍
Hybrid Retrieval
Dense (pgvector cosine) + Sparse (BM25 TF-IDF) results fused via Reciprocal Rank Fusion. Redis caches query results with configurable TTL for sub-millisecond repeat queries.
📊
RAGAS Evaluation
Faithfulness, Answer Relevancy, Context Recall, Context Precision — computed per sample with LLM-as-judge (OpenAI) or extractive fallback. All runs traced to LangSmith.
03 — Technology

TECH STACK

⚡
FastAPI
api/main.py
Backend
🐘
PostgreSQL 16
+ pgvector ext
Storage
🤗
BAAI/bge-large
1024-dim embeddings
ML Model
📈
BM25
rank-bm25 library
Retrieval
⚗️
RAGAS
evaluation framework
Eval
🔴
Redis 7
query + embedding cache
Cache
🔭
LangSmith
tracing + monitoring
Observability
☁️
GCP Cloud Run
serverless deployment
Infra
🏗️
Terraform
infrastructure as code
IaC
🔄
GitHub Actions
CI/CD pipeline
DevOps
🐳
Docker
containerization
Runtime
🧪
pytest
async test suite
Testing
04 — Interactive Demo

LIVE DEMO

Connected to the production API. Upload a document, query it with three retrieval strategies, and evaluate quality with RAGAS.

Document Ingestion

Drop a file or browse

PDF or TXT · up to 50 MB

512
uploading... 0%
—
doc_id: —
—
Chunks
—
Strategy
—
ms
Recent Queries
No queries yet — run one below
Query Configuration
pgvector cosine + BM25 → RRF fusion
5
⌕

Results appear here after a query

Side-by-Side Mode Comparison

Run the same query across all three retrieval strategies simultaneously and compare ranking differences.

Dense —
Run comparison to see results
Sparse —
Run comparison to see results
Hybrid (RRF) —
Run comparison to see results
RAGAS Evaluation

Evaluate retrieval quality with four RAGAS metrics. Add multiple samples for aggregate scoring.

05 — Reference

API ENDPOINTS

Full OpenAPI docs available at the live Swagger UI.

GET /health
Health check with dependency status for Postgres, Redis, and LangSmith.
statusstring"ok" when all services up
postgres_connectedbool
redis_connectedbool
POST /ingest/
Upload a PDF or TXT file. Returns document ID and chunk info.
fileFilePDF or plain text
chunking_strategystrfixed | semantic
chunk_sizeintcharacters per chunk
chunk_overlapintoverlap in characters
POST /query/
Retrieve relevant chunks using dense, sparse, or hybrid retrieval.
querystr1–1000 characters
retrieval_modestrdense | sparse | hybrid
top_kint1–50, default 5
score_thresholdfloat0.0–1.0
POST /evaluate/
Run RAGAS evaluation over live retrieval. Supports batch samples.
sampleslistquestion + optional ground_truth
retrieval_modestrdense | sparse | hybrid
top_kintchunks per sample