Fin-CX-Reranker Model
A specialized AI model that scores and ranks retrieved content for relevance, accuracy, and contextual fit. Evaluates each piece against the customer query, downranks outdated sources, and selects the final content for the LLM to use in response generation.
Why Reranking Matters
After the fin-cx-retrieval model identifies candidate content, the reranker model performs the critical task of determining which pieces are truly the best match. While retrieval casts a wide net to capture potentially relevant content, reranking applies precise judgment to select the optimal answer.
This two-stage approach balances recall (finding all relevant content) with precision (selecting only the best). The reranker model is specifically trained to make fine-grained distinctions between similar pieces of content, choosing the one that will produce the most accurate, helpful response.
The Reranking Process
The fin-cx-reranker model evaluates each retrieved candidate through multiple dimensions:
- Score relevance: Assigns precise relevance scores based on how well each piece matches the customer's specific query
- Evaluate context match: Considers conversation history, customer context, and situational factors to determine best fit
- Downrank outdated sources: Identifies and demotes content that may be stale, deprecated, or no longer current
- Output final selection: Provides the LLM with the highest-quality content for generating the customer response
Precision at Scale
The reranker model's specialized training on customer service data enables it to make nuanced decisions that generic models miss. It understands the difference between content that's technically related versus content that will actually resolve the customer's issue.
This precision is what enables the Fin AI Engine to achieve industry-leading accuracy rates. By ensuring the LLM works with only the most relevant, accurate content, the reranker model is a key component in preventing hallucinations and maintaining response quality at scale.