CiteWeave

Ι — Origin

How it started.

CiteWeave originated from research frustration. As an AI researcher, finding information is not difficult — understanding information is. Researchers constantly ask: what datasets are used? What metrics matter? What is novel? What is foundational? Existing systems retrieve information. But orientation remains expensive.

ΙΙ — The Problem

What we're solving.

Standard RAG is lazy. It takes a user's prompt, finds text chunks that share similar keywords (via vector embeddings), and tells the LLM to write an answer. In academia, law, and medicine, this is fatal. Standard RAG ignores the hierarchical importance of papers — it will treat a 1990 abandoned hypothesis with the same weight as a 2024 heavily-cited breakthrough simply because the keywords matched.

People do not want papers. People want context. People want relationships. People want neighborhoods. People want comparisons. People want significance. Literature should feel explorable, not exhausting.

ΙΙΙ — Differentiation

Why this, not that.

CiteWeave does not guess. It crawls the actual citation graph using PageRank-style algorithms to surface foundational impact — a 1990 heavily-cited paper will not be treated the same as a keyword match from last week. It physically extracts sentences from "Future Work" sections of papers and mathematically clusters them to reveal what a field does not yet know. No chatbot. No hallucinations. 100% verifiable, grounded in the structural hierarchy of science. The engine is inherently vertical-agnostic: swap the corpus and the LLM extraction prompts, and the same pipeline works for case law (CourtListener citations), clinical trials (PubMed), or regulatory compliance documents.

ΙV — Capabilities

What it does today.

• ›Citation graph traversal (PageRank-style)
• ›Comparable bucket clustering
• ›Novelty scoring
• ›Research gap identification
• ›Foundational & emerging work detection
• ›Dataset & metric extraction
• ›Cross-encoder reranking
• ›Methodology clustering
• ›2.5 TB local corpus (no API rate limits)
• ›Dissenting opinion extraction (legal vertical)

V — Target Audience

Who we're building for.

VI — Open Questions

What we don't know yet.

• How much synthesis is useful before it becomes trusted?
• Can machine-generated orientation become trusted in academia?
• Can research become navigable for non-experts?
• What domains most urgently need this?

Feedback wanted from: Researchers, academics, lawyers, scientists, policy professionals — anyone conducting literature reviews or dissatisfied with existing discovery systems.