What question did this study set out to answer?

This research aims to evaluate different retrieval strategies in a lightweight retrieval-augmented generation architecture using PostgreSQL.

July 7, 2026Open Access

Hybrid Retrieval-Augmented Generation on Serverless PostgreSQL: A Comparative Evaluation of HyDE, Graph-Based, and Fusion Retrieval Strategies

Key Points

This research aims to evaluate different retrieval strategies in a lightweight retrieval-augmented generation architecture using PostgreSQL.
Implemented a retrieval architecture utilizing PostgreSQL with pgvector and serverless edge functions.
Evaluated four retrieval strategies: hybrid search with RRF, HyDE, multi-query expansion, and graph-based retrieval.
Conducted evaluation on a 427-document Spanish-language corpus with 22 ground-truth pairs.
HyDE produced the highest point estimates with P@5: 0.595, MRR: 0.838, and a Hit Rate of 1.0.
The retrieval strategies were compared, but the small size of the ground-truth set (n=22) limits statistical strength.
Figures for Hybrid (RRF) and Graph-Based comparisons are preliminary estimates.

Abstract

We present a lightweight RAG architecture built entirely on PostgreSQL with pgvector and serverless edge functions, implementing and evaluating four retrieval strategies: hybrid search with Reciprocal Rank Fusion (RRF), Hypothetical Document Embeddings (HyDE), multi-query expansion, and graph-based retrieval over an automatically constructed knowledge graph. Evaluation on a 427-document Spanish-language corpus with 22 ground-truth pairs shows HyDE with the strongest point estimates (P@5: 0.595, MRR: 0.838, Hit Rate: 1.0). Because the ground-truth set is small (n=22) and no confidence intervals or significance tests are reported, these comparisons should be read as exploratory, and the Hybrid (RRF) and Graph-Based figures used for comparison are themselves preliminary estimates pending full cross-validated runs.

Read Full Paperexternally

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper