Back office ops · Production

Grab experiments with LLM-assisted vector similarity search for improved relevance on complex queries

The problem

Traditional vector similarity search struggles with nuanced and negation-based queries, returning semantically close but contextually wrong results, limiting its usefulness for complex search requirements.

First attempt

Raw vector similarity search failed to handle negations, returning dishes containing the explicitly excluded ingredients when queried for food with no fish or shrimp.

Workflow diagram · grounded in source
1
Vector search shortlisting
ai_action
“We first perform a vector similarity search on the dataset to obtain a shortlist of potential matches (e.g., top 10-50 results) for the given query. This step leverages the efficiency of vector similarity search to quickly narrow down th…”
2
LLM re-ranking
ai_action
“The shortlisted results from the vector similarity search are then fed into an LLM, which ranks the results based on their relevance to the original query. The LLM's ability to understand natural language queries and contextual informati…”
3
Top results output
output
“supply the shortlist and the original query to LLM with some descriptive instructions in the prompt to find the top three matches from the provided shortlist”
Reported outcome

The LLM-assisted approach showed a clear advantage over raw vector search for complex queries and was deployed internally at Grab for larger datasets with similar improvements in search result relevance.

Reported metrics
Search relevance for complex queriesclear advantage over the raw similarity search
search result relevance in internal Grab deploymentsimilar enhancement in search result relevance
Reported stack
FAISSGPT-4otext-embedding-ada-002
Source
https://engineering.grab.com/llm-assisted-vector-similarity-search
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The LLM-assisted approach showed a clear advantage over raw vector search for complex queries and was deployed internally at Grab for larger datasets with similar improvements in search result relevance.

What tools did this team use?

FAISS, GPT-4o, text-embedding-ada-002.

What results were reported?

Search relevance for complex queries: clear advantage over the raw similarity search; search result relevance in internal Grab deployment: similar enhancement in search result relevance (source-reported, not independently verified).

What failed first in this deployment?

Raw vector similarity search failed to handle negations, returning dishes containing the explicitly excluded ingredients when queried for food with no fish or shrimp.

How is this back office ops AI workflow structured?

Vector search shortlisting → LLM re-ranking → Top results output.

Related back office ops cases