Cui (formerly LitScribe) — Evidence-Grounded Literature Review and Research-Judgment Tool

Now: Cui (June 2026 – present)

Fluent LLM reviews mostly summarise and agree with the user, so Cui supports the researcher’s judgment instead of generating documents: question workspace → user-written claims → immutable review rounds in which an LLM challenges the claims → evidence marked supports / contradicts / silent / cannot-assess → human adjudication → landscape / gap view → related-work draft. Event-sourced (FastAPI, Postgres, React / TypeScript). Early development.

LitScribe v4 (January – June 2026)

LitScribe is an AI literature review engine that takes a research question and returns a 1,500–2,000 word review with LLM-checked citations, methodology comparison, research timeline, and BibTeX export — typically in about two minutes. Each [@key] citation is checked by an LLM against the source paper, and the system detects contradictions across papers and presents them as critical analysis.

Design Choices

1. Contradiction-aware synthesis — contradictions detected across papers are injected into the narrative and presented as critical analysis.
2. Metacognitive quality loop — after self-review, the system evaluates which pipeline steps failed and autonomously re-runs them with adjusted strategy. Strategy decisions are persisted to a knowledge base for future reviews.
3. Search-augmented refinement — when users say “add a section about X”, LitScribe searches for new papers on X first, analyzes them, then writes the section with citations from the newly found papers.

Architecture

DeepAgents Supervisor + Deterministic Pipeline

A natural-language supervisor (built on DeepAgents, 7 tools) interprets user intent, then dispatches a deterministic multi-stage pipeline with a metacognitive quality loop:

sequenceDiagram
    actor User
    participant S as Supervisor<br/>(DeepAgents)
    participant P as Pipeline

    User->>S: "Review CRISPR knockout strategies in CHO cells"
    S->>S: Understand intent + confirm
    S->>P: run_review(question, 12 papers)

    rect rgb(240, 248, 255)
        Note over P: Deterministic Pipeline
        P->>P: 1. Plan (decompose sub-topics)
        P->>P: 2. Search (6 academic databases + OA lookup)
        P->>P: 3. Read (analyze each paper)
        P->>P: 4. Contradictions (pairwise)
        P->>P: 5. Graph analysis (≥5 papers)
        P->>P: 6. Synthesize (parallel sections)
        P->>P: 7. Debate (reviewer ↔ synthesizer)
        P->>P: 8. Ground (LLM citation check)
        P->>P: 9. Review (self-review score)
    end

    alt Score < 0.65
        P->>P: Metacognition → re-run failed steps
    end

    P->>P: Save (session + knowledge base)
    P-->>S: Review + References + BibTeX
    S-->>User: 1,500-word review, score 0.82

Each numbered step persists to a SQLite knowledge base, so reviews can resume after interruption and downstream steps reuse upstream results across sessions.

Six-Source Academic Search

The Search step (step 2 in the pipeline) fans out across six databases in parallel, then funnels results through dedup → keyword filter → LLM selection → open-access PDF lookup:

graph TB
    Q[Research Question] --> PLAN[LLM Plan<br/>sub-topics + queries]
    PLAN --> |parallel| OA[OpenAlex<br/>250M+]
    PLAN --> |parallel| EPM[Europe PMC<br/>40M+]
    PLAN --> |parallel| PM[PubMed<br/>35M+]
    PLAN --> |parallel| S2[Semantic Scholar<br/>200M+]
    PLAN --> |parallel| AR[arXiv<br/>2M+]
    PLAN --> |parallel| CR[CrossRef<br/>140M+]

    OA --> DEDUP[DOI Dedup]
    EPM --> DEDUP
    PM --> DEDUP
    S2 --> DEDUP
    AR --> DEDUP
    CR --> DEDUP

    DEDUP --> KW[Keyword Filter]
    KW --> LLM[LLM Selection<br/>top N relevant]
    LLM --> UPW[Unpaywall<br/>OA PDF lookup]
    UPW --> OUT[Papers<br/>ready for analysis]

    style OA fill:#e8f5e9
    style EPM fill:#e8f5e9
    style PM fill:#e3f2fd
    style S2 fill:#e8f5e9
    style AR fill:#fff3e0
    style CR fill:#e8f5e9

Domain-aware routing auto-skips arXiv for biology/medicine; LLM-based selection filters irrelevant results after DOI deduplication.

Features

Citation Grounding

Every [@key] is checked by an LLM against the source paper; claims it judges unsupported are flagged and rewritten. LLM-checked citation-support rate: 62–100% across five benchmark queries.

Contradiction Detection

Pairwise comparison surfaces opposing findings, which are then injected into the review as critical commentary:

[@cho2025]: “FUT8 knockout had no negative effect on cell growth” [@lin2020]: “FUT8 knockout reduced cell viability by 40%” → classified as opposing_conclusions (major) → integrated into discussion

Multi-Agent Debate

After synthesis, a reviewer agent critiques the draft → synthesizer revises → up to 2 rounds. Catches unsupported claims, missing perspectives, and weak synthesis.

Local Paper Modes

• litscribe draft my_review.md paper1.pdf paper2.pdf — analyze your draft against references
• litscribe outline *.pdf refs.bib — “what review could I write from these?”
• litscribe augment "CRISPR delivery" papers/*.pdf — your papers + online search

Natural Language Interface

The supervisor understands intent in English or Chinese and routes to the right tool:

"写个关于transformer的综述，3个主题"     → run_review (Chinese)
"加一段关于delivery methods"             → refine_review (searches new papers first)
"这篇综述适合什么人看"                   → assess_reading_level

Cross-Lingual

Direct generation in target language (not post-translation), CJK-aware word counting, mismatch detection.

Security

HMAC integrity on the knowledge base to detect data poisoning, prompt injection filter, XSS prevention, rate limiting, PDF URL domain whitelist, parameterized SQL.

Review Output

Every generated review includes:

• Thematic analysis with [@key] Pandoc citations
• BibTeX-compatible reference list
• Methodology Comparison table across all papers
• Research Timeline (foundation → development → frontier)
• Statistical Summary — extracted p-values, effect sizes, sample sizes
• Suggested Figures — recommended visualizations with placement notes

Interfaces

• CLI — 12 commands (chat, review, draft, outline, augment, sessions, export, evaluate, serve, skills, …)
• Web UI + REST API — ~20 endpoints including SSE streaming for the review pipeline
• Sessions + Knowledge Base — persistent across reviews; supports diff-based refinement

Benchmark

LLM self-review score 0–1, one run per query (about ±0.10 between runs); citation support = % of citations an LLM judged supported by the source paper, not accuracy against human labels. Model: DeepSeek V4 Flash.

Technical Stack

Limitations & Roadmap

• Requires API keys for cloud LLM functionality
• PDF extraction quality varies with document formatting
• LLM-checked citation-support rate is lower in some CS / medicine queries (62–68%)
• The 49 unit tests do not cover the core pipeline, grounding, debate or search

LitScribe is frozen at v4.0.0; development continues as Cui.

GitHub Repositories

Cui · LitScribe (v4, frozen)

Timeline

LitScribe: January – June 2026 (frozen at v4.0.0)
Cui: June 2026 – present (early development)