Thalius AI
Thalius offers Hippocampus, a retrieval-augmented generation platform that retrieves the exact passage from a user's documents supporting each answer, explicitly stating when the material lacks the information. The system outperforms alternative retrieval tools in benchmark tests while lowering cost per query, and provides granular permission enforcement and full auditability. It supports flexible deployment across EU-hosted cloud, customer cloud, or fully on-premises environments.
- Artificial Intelligence
- Enterprise Software
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Standard retrieval-augmented generation (RAG) systems and vector search tools often return plausible but unverified answers, failing to pinpoint the exact source passage that supports a response. This leads to hallucinated or inaccurate outputs, especially on complex comparison and synthesis questions, while incurring high costs per query and lacking granular permission controls.
Solution
Hippocampus provides a retrieval-augmented generation platform that retrieves the exact passage from a user's documents supporting each answer, with explicit statements when the material does not contain the information. The system outperforms alternative retrieval tools in answer accuracy while lowering cost per query, as demonstrated in benchmark tests against OpenAI File Search, PageIndex, and vector stores. It offers flexible deployment options including EU-hosted cloud, customer cloud, or fully on-premises, with permission enforcement per user or agent applied at query time. Every answer records the documents that produced it, providing full auditability and verifiable outputs.
Target Audience
Primary customers are enterprises and teams that need verifiable, auditable answers from their own document repositories, particularly in regulated industries such as healthcare, life sciences, and legal sectors where answer accuracy and source citation are critical.
Features
- Exact-passage retrieval that cites the specific source document supporting each answer, with explicit refusal when documents lack the information
- Benchmark-verified accuracy: 40 of 48 correct answers on FDA document set versus 36 for OpenAI File Search and 15 for model alone; 92% success rate on FinanceBench and 86% on LegalBench-RAG
- Permission enforcement per user or agent, ensuring each entity sees only cleared content when the query runs
- Full auditability with every answer recording the documents that produced it
- Deployment flexibility across EU-hosted cloud, customer cloud, or fully on-premises with identical product functionality
- Scanned document ingestion with OCR capabilities, reading pages without text layers at ingest
- Model-agnostic design that works with multiple answering models, including GPT and Claude families
- Integration options including API, MCP, and web interfaces