01
Citation-first UX
The answer and its evidence were designed as one interaction, making verification a normal part of use.
CASE / 02 · Windrose product · research copilot
Causalith
Research Copilot
Our own research copilot turns papers into citation-grounded answers and connected knowledge, combining retrieval with a graph of the entities and claims inside the corpus.
RAGNeo4jVector embeddingsCitationsResearch
Grounded
Answers tied to retrieved passages
Connected
Entities and claims linked in a graph
Built in-house
Ingestion, retrieval, graph, interface
The challenge
General-purpose assistants can summarize quickly, but research work breaks when the source of a claim is unclear or the system loses the relationships between papers, entities, and ideas. Causalith needed to make a large corpus searchable without flattening it into anonymous chunks—and it needed to keep the evidence visible inside every useful answer.
What we built
We built the ingestion pipeline, retrieval layer, Neo4j graph construction, answer generation, citation experience, and product interface in-house. Papers are parsed, cleaned, chunked, and embedded; entities and claims are lifted into a graph; and each question retrieves the most relevant passages and connected context before an answer is generated. The interface keeps citations close enough to inspect instead of hiding provenance behind a generic response.
IngestionDocument parsing, cleanup, chunking, metadata, embeddings, and repeatable corpus updates.
RetrievalSemantic search and ranking over passages selected around the research question.
GraphEntity and claim extraction into Neo4j so connected ideas survive beyond individual chunks.
AnswerGrounded generation with passage-level citations and a product experience designed for verification.
A research product built around traceable answers.
Causalith demonstrates the Windrose approach to production AI: retrieval quality, graph context, citations, and interface decisions are treated as one product system. It is live at causalith.com ↗.
/02 — Engineering decisions
The choices beneath the interface.
The strongest AI product decisions are often invisible to the user. These were the boundaries that kept the system coherent, maintainable, and ready for real use.
02
Hybrid context
Vector retrieval handles semantic relevance while graph context preserves useful relationships across the corpus.
03
Visible provenance
The system exposes source passages instead of asking users to trust a confident paragraph with no trail.
04
Evaluable pipeline
Ingestion, retrieval, graph enrichment, and generation remain separate enough to diagnose and improve.
/03 — Continue exploring
Related work + capabilities.
Follow the system into the service behind it, or compare it with another product Windrose has shipped.
Ready to make it
real?
Send the problem, the context, and the honest budget. We will reply within one business day with useful next steps—even when the right answer is not us yet.
hello@windroselabs.com ↗