Enterprise APIThe AI engine behind the #1 AI agent for customer service.
Now available as an API.

Incoming queries are optimized before they hit the retrieval pipeline. Context from the full conversation history, user metadata, and timestamp are distilled into a semantically rich search query — maximizing retrieval precision from the first hop.

Our proprietary fin-cx-retrieval model — an embedding model trained on signal extracted from millions of Fin retrievals — scans your knowledge base and returns the top 40 candidate documents by semantic similarity.

Training uses InfoNCE contrastive loss with hard positives and hard negatives mined from production query logs — teaching the model the nuances of customer support language across industries and languages.

The fin-cx-reranker has an 8,192 token context window, using ModernBERT-large as a building block. It reorders the 40 candidates by true relevance, trained using RankNet loss across 10s of millions of passage pairs.

Replaces other commercial reranking services with lower latency, higher throughput, and significantly higher quality on customer service content.

A multi-task ModernBERT classifier making three simultaneous predictions: escalation decision (3-way classification), reason categorization, and guideline citation (multi-label). Trained on millions of multilingual examples.

Achieves >98% accuracy at a competitive cost and a fraction of the latency of larger models often used for this task.

The top-ranked documents, full conversation context, user metadata, and brand guidelines are synthesized into a precise, cited response. A dedicated validation layer then checks for accuracy, hallucination, tone alignment, and policy compliance before the response is returned. Every answer is grounded in your content — never fabricated.