Research
Analyze papers and reports with multi-step reasoning. Every answer is cited back to the source.
What you can do
Deep Multi-Step Queries
Complex questions are broken down and answered across multiple documents automatically.
Automatic Relationship Mapping
Discovers connections between concepts, people, and findings across your research library.
Source-Cited Answers
Every response includes inline citations linking to specific documents and pages.
Conflict Detection
When sources disagree, discrepancies are flagged and attributed to each document.
Relevant templates
Frequently asked questions
What is multi-hop retrieval and when does it activate?
Multi-hop retrieval performs multiple rounds of search with query expansion and HyDE (Hypothetical Document Embeddings) to answer complex or multi-faceted research queries. The QueryRouter classifies incoming queries and routes complex, vague, or analytical questions to the multi-hop strategy automatically. Within each hop, all search queries (original, expansion, and HyDE) run concurrently via asyncio.gather for speed. Simple factual queries use a single retrieval pass.
How does the knowledge graph improve search results?
During document upload, the NER model extracts entities and the knowledge graph builds relationships between them using co-occurrence analysis with PMI (pointwise mutual information) weighting. At query time, Graph-RAG expansion uses these entity relationships to find related context that keyword or semantic search alone would miss. For example, if you query about a specific protein, the graph can surface related genes, pathways, and diseases from across your corpus. The graph uses a dual-store architecture: PostgreSQL as the source of truth and Neo4j for graph traversal queries.
How are citations tracked in RAG responses?
Every RAG response includes inline citations with source attribution. The CitationExtractor performs title normalization and numeric extraction to link response claims to specific retrieved chunks. Citations reference the document title, page number, and relevant snippet. In multi-document scenarios, the context builder groups chunks by document with clear headers, and the prompt instructs the model to attribute each claim to its source document by name.
Start using Leepi.ai for research
Upload your documents and start asking questions. No credit card required.
Start Free