The Cognir Ontology™

Raw Text Understanding

The system performs foundational semantic extraction on the user's raw input. It does not generate questions yet. It does not infer beyond the text. It extracts exactly what is present: the core problem, the apparent goal, key concepts, research domains, and notable themes.

{
  "main_problem": "string describing the core problem",
  "main_goal": "string describing what the user wants to understand",
  "key_concepts": ["concept1", "concept2", "concept3"],
  "research_domains": ["domain1", "domain2"],
  "notable_themes": ["theme1", "theme2"]
}

Intent Compression

This is the critical interpretive layer. The system identifies what the user is actually trying to understand — not what they said, but what they meant. It detects the knowledge gap they are circling around and articulates the high-level research objective that would genuinely serve them.

Constraint: All inference is grounded in the provided extraction. No hallucinated domains. No invented concepts. The system is explicitly prohibited from adding external knowledge.

Candidate Question Generation

Generates exactly 10 candidate research questions derived solely from the compressed intent and extracted concepts. Questions are varied in approach — some empirical, some theoretical, some applied — but all are researchable, specific, and traceable to the user's original input.

{
  "questions": [
    "How does X mediate the relationship between Y and Z in population P?",
    "What is the comparative efficacy of A versus B under condition C?",
    "..."
  ]
}

Search Query Generation

For each candidate question, the system generates 3 targeted academic search queries. These are not the full question text. They are short (3-7 words), academic-focused, and varied in angle. The system uses Google Scholar operators: intitle:, author:, source:, quoted phrases, and boolean OR.

Evidence Collection

Executes Serper searches for each query set, then crawls priority academic domains with concurrency control. The crawler extracts title, abstract, headings, body text, publication date, and keywords. Priority domains are queried first: arXiv, Nature, Springer, IEEE, ACM, PubMed, NIH, ResearchGate, Frontiers, ScienceDirect, JSTOR, and others.

Resilience: If a crawl fails, the pipeline continues. Evidence is additive, not blocking. The system degrades gracefully.

Viability Scoring

Each candidate question receives a composite viability score (0-100) computed across five weighted dimensions. This is deterministic scoring, not LLM inference. It runs fast and produces reproducible results.

LLM Validation Pass

The top 5 candidates are submitted to a second LLM pass for evidence-backed validation. The model evaluates whether the question is active (ongoing research exists), viable (answerable with current methods), and researchable (empirically or theoretically investigable). It does not hallucinate — it bases its judgment on the compressed evidence summary provided.

{
  "is_active": true,
  "is_viable": true,
  "is_researchable": true,
  "evidence_assessment": "brief assessment",
  "confidence": 0.0 to 1.0,
  "key_finding": "most important finding"
}

Question Improvement

Each validated question is refined for specificity and measurability. The system preserves the exact original topic and intent — it does not change the subject matter. It adds specificity (who, where, what population, what context) and makes the question measurable where possible. If improvement fails, the original question is retained.

Ranking & Selection

Final composite ranking combines the deterministic viability score with validation bonuses (active +10, viable +10, researchable +5) and confidence-weighted adjustments. The top 3 questions are selected and presented with full provenance: original draft, improved version, evidence sources, scoring breakdown, and a direct handoff to Phase 2.

Question Decomposition

The system decomposes the research question into structured components for systematic literature search. This includes domain classification, MeSH term extraction, sub-question generation, population/exposure/outcome identification, confounder detection, and controversy mapping.

{
  "domain": "primary scientific domain",
  "summary": "1-2 sentence restatement",
  "key_concepts": ["concept1", ...],
  "mesh_terms": ["MeSH term 1", ...],
  "sub_questions": ["specific sub-question 1", ...],
  "population": "who/what is studied",
  "exposure_intervention": "what is being tested",
  "outcome": "what is being measured",
  "confounders": ["potential confounder 1", ...],
  "known_controversies": ["controversy 1", ...],
  "inclusion_criteria": "what papers to include",
  "exclusion_criteria": "what papers to exclude",
  "sparse_areas": ["underrepresented topics"],
  "suggested_journals": ["journal name 1", ...]
}

Search Strategy Planning

The system generates 14-20 varied search queries across 10 strategic categories. This is not brute force. It is surgical precision: primary queries for direct hits, secondary queries for foundational papers, review queries for meta-analyses, methodology queries for study design, author queries for key researchers, date-range queries for recency or classics, synonym expansions for coverage, grey literature queries for reports, and preprint queries for cutting-edge work.

Multi-Query Discovery

Executes all planned queries against Serper with intelligent caching. Results are deduplicated in real-time. Each query returns up to 10 organic results. The system respects rate limits with 250ms delays between calls and serves cached responses when available (48-hour TTL).

Deduplication & Filtering

A three-layer deduplication pipeline: (1) Fast fuzzy title matching in JavaScript with 82% similarity threshold and prefix hashing. (2) Academic domain filtering against an expanded whitelist of 25+ domains and title pattern matching. (3) LLM semantic deduplication for near-duplicates and off-topic removal on the top 45 candidates.

Metadata Enrichment

The top 40 deduplicated candidates are enriched via three free academic APIs in priority order: Semantic Scholar (abstracts, citations, venues, PDF links), CrossRef (DOI resolution, publication dates, author lists), and arXiv (preprint abstracts, submission dates). Title similarity matching ensures the correct paper is retrieved even when exact titles differ slightly.

Relevance Scoring

Enriched papers are scored in batches of 12 via LLM evaluation. The model receives the research question, decomposition, and full enriched metadata (abstract, year, venue, citations, authors). It returns relevance (0-100), quality signal, paper type, category, priority, and a 1-2 sentence justification. Papers scoring below 40 are excluded.

Citation Snowballing

The top 6 must-read and high-relevance papers become seeds for recursive discovery. The system generates snowball queries targeting citations, references, and co-author networks. Novel candidates are deduplicated against existing results, enriched, and scored. This closes the coverage gap that initial queries inevitably leave open.

Reading List Curation

The final curation stage structures the scored papers into a logical reading sequence. A senior-researcher LLM persona evaluates the full pool and produces: an overview synthesis, key themes, a 5-10 paper reading order, a single "start here" recommendation, and six taxonomic categories. The system is ruthless: maximum 25 papers total. Every paper must earn its place.

Gap Analysis & Synthesis

The final stage identifies what is missing. A methodologist LLM evaluates the curated list against the original decomposition, flags sparse areas, and suggests next search strategies. It reports coverage confidence (high/medium/low) with specific reasoning. This is not a generic disclaimer — it is a targeted assessment of whether the current list adequately covers the research question.

{
  "gaps": [
    {"area": "gap topic", "description": "what's missing and why it matters"}
  ],
  "coverage_summary": "1 sentence on what the list covers well",
  "next_steps": ["suggested next search strategy 1"],
  "confidence": "high|medium|low",
  "confidence_reason": "brief reason"
}