Research & Evaluations
1 min read
Introduction
This section shares two kinds of work done at Unique:
Research Exploratory initiatives that serve as a foundation for future product innovations. They are often experimental or work in progress, and show the questions we are investigating and the approaches we are testing.
Evaluation Repeatable benchmarks of language models, competing products and our own features. Each one describes its setup and data, so results can be compared over time.
The goal is not polished documentation, but transparency into our thought process, prototypes, measurements and learnings. Many of these efforts eventually influence or evolve into features of our production systems. Results reflect the models and versions available when each piece of work was done.
Contents
Model and Chat Benchmarks
How language models and reasoning-effort settings perform across our agents, and how well Unique AI chat answers from internal documents and the web.
Evaluation Models and Harness Eval Overview
Evaluation Unique AI Chat Benchmarks
Competitive Benchmarks
How Unique AI compares with Microsoft Copilot and ChatGPT on complex financial and regulatory documents. The summary page links to the full evaluation data.
Ingestion, Retrieval & RAG
From document ingestion through search and retrieval to answer generation.
Overview Ingestion, Retrieval & RAG
Research Graph-based RAG Research
Evaluation RAG Assessment and Improvement
Evaluation Improving web search quality by switching to Brave or Perplexity
Agents & Tool Use
How our agents choose and use tools, and how well they complete end-to-end tasks with them.
Overview Agents & Tool Use
Evaluation Agentic Table Benchmark Reports
Evaluation Client-Ready Decks: Benchmarks
Hallucination
How reliably our hallucination check detects answers and files that contradict their sources, and which judge models work best.
Evaluation Remy Hallucination Benchmarks