K AVAILABLE
OPEN TO WORK
← BACK TO PROJECT

PROJECT

Enterprise RAG & Empirical Evaluation Pipeline

Enterprise-grade Retrieval-Augmented Generation system with Hybrid Search (Dense + BM25), neural cross-encoder reranking, and continuous RAGAS evaluation.

2026AI & Backend EngineershippedRAG & Information Retrieval

SYSTEM / ARCHITECTURE

How it fits together

  1. 01Hierarchical and semantic chunking engine preserving parent document context.
  2. 02Parallel hybrid retrieval combining Qdrant dense vector embeddings with BM25 sparse keyword search.
  3. 03Reciprocal Rank Fusion (RRF) algorithm merging heterogeneous retrieval distributions.
  4. 04Neural cross-encoder reranking stage filtering out irrelevant distractor chunks.
  5. 05Automated evaluation harness tracking Faithfulness, Context Recall, and Answer Relevance.
Faithfulness Score
89.4%
Context Recall
91.2%
Inference Cost
-55% Token Waste

Overview

Naive RAG pipelines (chunking text naively into 500-token blocks and performing single-vector cosine similarity) inevitably break down in enterprise settings. They suffer from context fragmentation, high false-positive retrieval on technical jargon, and vector hallucination without verifiable attribution.

This project implements an Enterprise RAG & Evaluation Pipeline featuring hybrid retrieval (Dense Vector + BM25 Sparse Search), cross-encoder neural reranking, and continuous empirical benchmarking using the RAGAS framework.

---

🏛 System Architecture

The pipeline is engineered for high accuracy, grounded responses, and verifiable attribution:

  • Multi-Strategy Ingestion Engine — Chunks long-form technical manuals, API specs, and domain PDFs using semantic boundary splitting and hierarchical parent-document referencing.
  • Hybrid Retrieval Layer — Concurrently executes dense semantic retrieval via Qdrant / pgvector alongside sparse keyword scoring via BM25.
  • Reciprocal Rank Fusion (RRF) — Merges dense and sparse score distributions into a calibrated candidate ranking.
  • Cross-Encoder Neural Reranker — Applies a secondary BGE / Cohere cross-encoder pass on top-$K$ candidates to score true query-passage semantic relevance, filtering out distractor chunks.
  • Context-Constrained Generator — Prompts decoupled LLM providers with strict citation enforcement, returning generated responses paired with chunk bounding IDs.
  • Automated RAGAS Evaluation Harness — Computes ground-truth metrics for Faithfulness, Answer Relevance, Context Recall, and Context Precision.

---

⚙️ Key Technical Challenges & Solutions

  • The Jargon & Acronym Problem — Dense embeddings often fail to distinguish specialized acronyms or serial numbers. Pairing dense embeddings with BM25 sparse search ensures exact identifier matches while preserving conceptual semantic search.
  • Token Efficiency & Noise Reduction — Passing 10 raw chunks to an LLM dilutes attention and increases latency. The cross-encoder reranker trims candidate sets down to the 3 most relevant passages, cutting prompt token costs by over 55%.
  • Hallucination Prevention — Enforcing strict markdown citation tags tied to vector database chunk IDs ensures every assertion is traceable to source documentation.

---

📊 Evaluation & Empirical Benchmarks

Benchmarked across an enterprise corpus of 500+ complex technical documents:

  • 89.4% Faithfulness Score — Verifiable factuality of model answers against retrieved context.
  • 91.2% Context Recall — Accurate retrieval of all required source passages for complex multi-hop queries.
  • 118ms Average Retrieval Latency — Optimized hybrid search execution with asynchronous pipeline dispatch.
  • 55% Reduction in LLM Inference Costs — Achieved through precision reranking and context compression.
TOPICS:
RAGVector SearchHybrid RetrievalRAGASQdrantPythonCross-Encoder