Customer support · Production

Assembled improves RAG results with Hybrid Search and Reciprocal Rank Fusion

The problem

Vector-only search failed to return relevant results for specific keyword queries in customer support, particularly when knowledge bases were uncurated and queries contained short or ambiguous keywords.

First attempt

A weighting-based fusion approach was tried first but proved unworkable because similarity score distributions varied widely across customers, making it impossible to determine universal weights.

Workflow diagram · grounded in source

Support query submitted

trigger

“designed to assist customer support by suggesting potential answers to support queries”

Parallel vector and keyword search

ai_action

“searching for a document could be done in parallel across multiple document stores using a standardized query”

Reciprocal Rank Fusion

ai_action

“Rank assignment: Each document from the individual ranked lists is assigned a score based on its rank position. Typically, the score is the reciprocal of its rank (i.e., 1/rank).”

RAG answer generation

output

“We use Retrieval Augmented Generation (RAG) for much of this pipeline”

Document store sync

integration

“Our PostgreSQL database tracks the synchronization status of each store. If a store fails to sync, we employ exponential backoff to retry the operation”

Reported outcome

Hybrid search combining vector and keyword search, merged via Reciprocal Rank Fusion, consistently outperformed more complex methods and enhanced the accuracy and relevance of search outcomes across a diverse customer base.

Reported metrics

Search result accuracy and relevanceenhanced the accuracy and relevance of search outcomes

RRF vs other fusion methodsconsistently outperformed many of the more complex methods we evaluated

Reported stack

Retrieval Augmented Generation (RAG)PineconeAlgoliaPostgreSQLS3

Source

https://www.assembled.com/blog/better-rag-results-with-reciprocal-rank-fusion-and-hybrid-search

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Hybrid search combining vector and keyword search, merged via Reciprocal Rank Fusion, consistently outperformed more complex methods and enhanced the accuracy and relevance of search outcomes across a diverse customer…

What tools did this team use?

Retrieval Augmented Generation (RAG), Pinecone, Algolia, PostgreSQL, S3.

What results were reported?

Search result accuracy and relevance: enhanced the accuracy and relevance of search outcomes; RRF vs other fusion methods: consistently outperformed many of the more complex methods we evaluated (source-reported, not independently verified).

What failed first in this deployment?