Comcast, like many other companies, provides customer service through various communication channels. While digital automation capabilities have been replacing human customer representatives for many tasks, there are still many situations that require human-to-human interactions.
To assist agents in these complex scenarios, Comcast has introduced the "Ask Me Anything" (AMA) feature. AMA allows agents to ask questions to a large language model (LLM) on demand, as they are handling customer conversations. The LLM provides accurate responses in real-time, reducing the amount of context switching the agent needs.
The key aspects of the AMA system are:
Document Preprocessing: Comcast standardizes documents into plain text, chunks them into snippets, and assigns unique identifiers to enable efficient retrieval.
Retrieving Relevant Text Snippets: Comcast experimented with various retrieval models, including sparse (BM25) and dense (DPR, MPNet-base) approaches. They found that OpenAI's ada-002 embeddings performed best on their evaluation set.
Reranking Search Results: Comcast used synthetic data generation and fine-tuning to improve the relevancy of search results, leading to better answer quality.
Generating Answers from Snippets: Comcast follows the conventional Retrieval-Augmented Generation (RAG) approach, where the LLM is prompted to generate answers using the retrieved snippets. They also implemented a "Citation Rail" to provide references for the answers.
Offline Response Evaluation: Comcast evaluated the system's responses using human-annotated ground truth answers and metrics like Answer Quality, Citation Match Rate, and Recall@K.
In internal experiments, Comcast found that agents using AMA versus a traditional search experience spent approximately 10% fewer seconds per conversation containing a search, translating to millions of dollars of savings annually. Agents also provided positive feedback nearly 80% of the time, demonstrating the usefulness of the AMA feature.
Comcast further conducted an A/B test of the reranker module, observing a statistically significant decrease in the "No Answer Rate" and an increase in the "Positive Feedback Rate" for the treatment group using the reranker.
Overall, Comcast's AMA feature showcases how large language models can be effectively integrated into customer service workflows to enhance agent productivity and customer satisfaction.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Scott Rome,T... at arxiv.org 05-03-2024
https://arxiv.org/pdf/2405.00801.pdfDeeper Inquiries