Customer support · Production

Thomson Reuters Labs builds RAG-powered customer support assistant using GPT-4 and milvus to reduce resolution times

The problem

Customer support agents at Thomson Reuters faced cognitive overload navigating hundreds of thousands of knowledge base articles and CRM tools, while resolution knowledge was siloed in individual agents with no structured way to share it across the team.

Workflow diagram · grounded in source
1
Ingest and chunk source data
integration
“we take the data from Knowledge base articles, CRM tools etc. and process these text files to chunks”
2
Generate and index embeddings
ai_action
“we used milvus which is an open-source vector database and used an opensource sentence transformer from huggingface `all-MiniLM-L6-v2`”
3
Agent submits query
trigger
“build a chatty interface for our customer support agents to get the most relevant solution from our curated database”
4
Dense retrieval of relevant docs
ai_action
“When a query is input, it is also encoded into a dense vector. The similarity between the query vector and the document vectors is computed using various distance metrics such as cosine similarity or Euclidean distance.”
5
GPT-4 generates grounded response
ai_action
“We used OpenAI's GPT-4 API. So, the most relevant context is concatenated along with the prompts and is sent to the API to get appropriate response for the user”
6
Accurate resolution delivered
output
“This response is most accurate as this match with the resolution to solve the issue within our products with the most recent information”
Reported outcome

Thomson Reuters Labs deployed a RAG-based chatty interface that provides customer support agents with accurate, product-specific resolutions grounded in curated domain knowledge, reducing resolution times compared to an ungrounded GPT-4 response.

Reported metrics
Resolution timereduce resolution times
Reported stack
GPT-4milvusall-MiniLM-L6-v2huggingface
Source
https://medium.com/tr-labs-ml-engineering-blog/better-customer-support-using-retrieval-augmented-generation-rag-at-thomson-reuters-4d140a6044c3
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Thomson Reuters Labs deployed a RAG-based chatty interface that provides customer support agents with accurate, product-specific resolutions grounded in curated domain knowledge, reducing resolution times compared to…

What tools did this team use?

GPT-4, milvus, all-MiniLM-L6-v2, huggingface.

What results were reported?

Resolution time: reduce resolution times (source-reported, not independently verified).

How is this customer support AI workflow structured?

Ingest and chunk source data → Generate and index embeddings → Agent submits query → Dense retrieval of relevant docs → GPT-4 generates grounded response → Accurate resolution delivered.

Related customer support cases