How it works
Three steps. That's it.
No engineering needed. No configuration required.
Drag and drop your PDFs, Word files, presentations. DEPOSIUM® reads, analyzes and structures them automatically. Even scanned documents (built-in OCR).
In English, in French, like you'd talk to a colleague. "What's the maintenance procedure for the compressor?" — DEPOSIUM® searches across all your documents and answers with sources.
DEPOSIUM® doesn't just search: it maps the links between your documents. Cross-references, contradictions, recurring themes — your knowledge takes shape.
From your files to your answers
From ETL to ET(K)L
Your documents, regardless of the source (PDF, DOCX, databases, APIs)
Text extraction, OCR, intelligent chunking, cleaning
Enrichment at ingestion: knowledge graph, contextual ontology, structural analysis, temporality
Indexing into a hybrid engine (BM25 + vector + graph) ready to be queried
Because we enrich at ingestion, every query is more precise. No costly post-processing, no hallucination from lack of context.
Under the hood: hybrid 5-signal search
Vector search alone loses precision as volume grows. DEPOSIUM® combines 5 independent mechanisms to maintain quality at scale.
Why standard RAG systems hallucinate — even with few documents
The problem isn't volume. Even with 10 documents, a standard RAG system can fabricate answers. The AI retrieves passages that are "close" to your question, but not necessarily the right ones — and generates a fluent answer regardless. The worst part: you can't tell the difference.
Meaning-based semantic search (not just keywords) on a quantized vector index — compact (~8× lighter) and high-recall, even at scale.
Exact keyword search. Finds precise references (part numbers, proper names, acronyms) that vector search may miss.
Traverses entity relationships to find indirectly related documents. Vector search never makes these connections.
Evaluation via Reciprocal Rank Fusion, false positives neutralized.
The final judge actually re-reads the question and each candidate document.
Hallucination rate by document volume
↓ Lower is better
| Error source | Standard RAG * | DEPOSIUM® |
|---|---|---|
| Wrong document retrieved | 1 to 15% | < 1% (5 signals + reranker) |
| Answer fabricated by the LLM | 3 to 10% | 0% measured (mandatory source citation) |
| Incomplete context (info spread across docs) | 5 to 15% | < 2% (knowledge graph) |
| Estimated overall hallucination rate | 10 to 30% | 0% measured on 800+ pages |
* Standard RAG estimates based on industry benchmarks (RAGAS, faithfulness). DEPOSIUM®: measured on internal 800+ page benchmark.
Measured performance, not promises
Benchmark over 10,000 queries on the knowledge graph engine *, local and cloud versions.
| DEPOSIUM® Local | DEPOSIUM® Cloud | FalkorDB | Neo4j | |
|---|---|---|---|---|
| Response time | 2.15 ms | 3.27 ms | 12.8 ms | 164 ms |
| Memory used | 147 MB | 136 MB | 850 MB | 4,200 MB |
| Throughput (req/s) | 1,480 | 545 | 15,000 | 2,800 |
* The knowledge graph engine is one of DEPOSIUM®'s 5 hybrid search signals (BM25, semantic, graph, RRF fusion, reranking). These signals are combined to produce every chat answer. Benchmark on Intel i5-10310U, 16 GB RAM, Linux.
Tech stack
Compliant by design.
| Layer | What we use | Why |
|---|---|---|
| Interface | Modern web application | Fast, works on any device |
| Analysis | Knowledge graph | Understands relationships between your documents |
| AI Models | Open-source, from 250 MB to 120B | Precise, auditable, lightweight |
| Security | Role-based access control | Fine-grained permissions for your team |
| Storage | S3-compatible or local | Your files stay where you decide |
| Integrations | MCP server, REST API, CLI | Plugs into your existing tools |
| Deployment | Docker / Kubernetes | European cloud, on-premise, or hybrid |
Ready to unlock what your documents know?
Free trial, no credit card. Your first results in minutes.