Back to Home

Case Study

Research Paper Discovery & Semantic Search

AI-Powered Scientific Literature Search Engine

PythonFastAPIPostgreSQLFAISSSentence TransformersMistral 7BRedisSQLAlchemy
Deployed
GitHub

The Problem

Challenge & Background

Researchers struggle to discover relevant scientific literature across vast corpora. Keyword searches miss semantically related papers that use different terminology.

The Solution

System Approach

Built an end-to-end paper discovery platform combining arXiv scraping, PostgreSQL metadata storage, and 384-dimensional Sentence-Transformers embeddings stored in a FAISS index with Mistral 7B structured insights.

Architecture

System Architecture & Data Flow

arXiv Scraper

Paper ingestion

PostgreSQL

Relational storage

MiniLM-L6

384-dim embeddings

FAISS Index

Vector search

FastAPI

REST endpoints

Redis Cache

Query caching

Mistral 7B

Paper analysis

Search Results

Ranked output

Engineering Decisions

Trade-offs & Technical Choices

Why hybrid FAISS + PostgreSQL architecture?

PostgreSQL provides structured filtering (authors, categories, publication dates) while FAISS handles high-dimensional vector similarity — isolating search and metadata workloads.

Why Redis caching layer?

Semantic embeddings and vector similarity lookups are computationally intensive. Redis caching delivers sub-500ms response times for repeated queries.

Why MiniLM-L6 embeddings?

384-dimensional dense vectors achieve exceptional semantic similarity performance with minimal memory footprint and fast inference.

Capabilities

Key Features & Engineering Scope

  • Automated arXiv paper ingestion and parsing
  • Semantic similarity search via FAISS
  • 384-dimensional Sentence-Transformer embeddings
  • FastAPI REST backend with Pydantic schemas
  • Redis caching for sub-500ms repeated queries
  • Batch paper insight extraction via Mistral 7B
  • Relational schema indexing 13,000+ papers

Outcomes

Results & Impact

13,000+

Research Papers Indexed

384-dim

Vector Embeddings

<500ms

Cached Query Latency

Full-stack architecture indexing 13,000+ research papers with sub-500ms cached query performance and automated LLM insights.