Research begins in chaos. A researcher does not wake up with a perfectly formed hypothesis — they wake up with a hunch, a contradiction, a pattern noticed in passing, or a frustration with existing literature. The gap between that raw cognitive state and a rigorous, defensible research question is where most projects die.
Traditional tools treat research as a search problem. You type a query, you get papers. But the researcher does not yet know what to query. They do not yet have the vocabulary. They have not yet articulated the boundary between what they know and what they need to know.
The COGNIR ONTOLOGY™ treats research as a transformation problem. It accepts unstructured, half-formed, emotionally charged human thought as input. It outputs ranked, evidence-grounded research questions and a curated literature pathway. The messy stage of research — the stage where most people quit — is compressed from weeks to hours.
9
Pipeline Stages
Per phase, fully autonomous
3
Enrichment APIs
Semantic Scholar, CrossRef, arXiv
∞
Query Variations
Synonym expansion + snowballing
Parses unstructured, stream-of-consciousness researcher notes into structured semantic components: core problem, knowledge gap, key concepts, research domains, and notable themes.
Generates 10+ candidate research questions derived solely from extracted intent. No hallucination. No external injection. Every question is traceable to the user's original input.
Executes multi-query Serper searches, crawls priority academic domains (arXiv, Nature, PubMed, IEEE), and extracts structured metadata including abstracts, publication dates, and citation counts.
Multi-dimensional scoring across five axes: Research Activity, Academic Coverage, Specificity, Novelty, and Practicality. Weighted composite produces a final 0-100 viability score.
Organizes discovered papers into six taxonomic categories: Foundations, Core Evidence, Frontiers, Methodology, Reviews & Meta-Analyses, and Controversies. Each paper is tagged with relevance score and reading priority.
Recursively searches citations and references of top-scored papers to discover seminal works and recent developments that initial queries may have missed.
The first engine accepts raw researcher cognition — notes, ramblings, half-formed hypotheses — and transforms it into a ranked set of 3 validated research questions. This is not keyword extraction. It is semantic archaeology: digging beneath the surface text to find what the researcher actually means.
Early Access
The documentation is comprehensive. The system is more so. Request access to experience the full pipeline on your own research.
The second engine accepts a refined research question (from Phase 1 or direct input) and produces a structured, categorized reading list with full provenance. It does not just find papers. It understands the topology of a research field and maps the user's position within it.
The system routes all LLM calls through OpenRouter, enabling multi-key rotation for resilience. Four API keys are maintained in a round-robin pool with automatic failover. If one key exhausts its rate limit or fails, the next key is attempted immediately. After two full passes through the pool, the system backs off with exponential delay.
Poolside Laguna M.1 (Phase 1) GPT-OSS 120B (Phase 2) Temperature: 0.15-0.2 Max Tokens: 900-2400
Google Search API via Serper.dev. Returns organic results with title, snippet, URL, and position. Supports up to 10 results per query. All responses are cached locally for 48 hours to minimize API usage and improve latency on repeated topics.
Three free, no-key academic APIs provide metadata enrichment. Each has a 168-hour (7-day) cache TTL. Title similarity matching prevents false positives when exact titles differ.
Semantic Scholar
graph/v1/paper/search
arXiv
export.arxiv.org/api
Uses allorigins.win CORS proxy for cross-origin page fetching. DOMParser extracts structured content: title, meta description, abstract selectors, headings (h1-h3), body text (max 1500 chars), publication dates, and keywords. Noise elements (scripts, nav, ads, sidebars) are stripped before extraction. Priority domain sorting ensures academic sources are crawled first.
Research inputs are processed in real-time and never stored on Cognir servers. All caching is local to the user's browser via localStorage. No training data is collected from user queries.
Every output is traceable to either the user's input or retrieved evidence. The system is explicitly instructed to not infer beyond provided text. When evidence is insufficient, the system reports low confidence rather than inventing sources.
OpenRouter keys are rotated automatically with exponential backoff. No single key bears full load. Failed keys are logged but never exposed to the user interface. The system degrades to partial results rather than failing entirely.
The system does not write original research, fabricate data, or generate citations that do not exist. It is a discovery and curation tool, not a content generator. All paper links are direct to source publishers or preprint servers.
Q3 2026 — Private Beta
200 researchers. Full two-phase pipeline. Export to Zotero, Mendeley, and BibTeX.
Q4 2026 — Collaborative Workspaces
Shared research projects, annotation layers, advisor review workflows, institutional licenses.
Q1 2027 — Live Literature Monitoring
Automated alerts for new papers matching your research questions. Weekly digest of frontier developments.
Q2 2027 — Causal Inference Layer
Automated identification of causal claims, confounder analysis, and study design quality assessment.
Early Access
You now understand exactly what the system does, how it does it, and why it is built this way. The only thing left is to use it. We are accepting 200 researchers for the private beta. If you are serious about your research, this is where you request access.
No commitment. No credit card. Just research.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。