Scientific literature review & research assistant

Millions of papers appear each year; no researcher, R&D team or student can read the relevant part of them. AI research assistants search scholarly indexes, screen candidates against stated criteria, extract data from full texts and write cited syntheses. They shorten the path from question to a defensible overview, but every claim must remain traceable to a source, and hallucinated references remain the central failure mode.

Problem & value

• Today: keyword searches in several databases, manual title/abstract screening, reading PDFs, copying data into spreadsheets, writing the synthesis. A systematic review is slow and labour-intensive; quick scoping for a product, grant or thesis is often skipped or shallow.
• AI value: broader recall through semantic and citation-graph search, assisted screening and data extraction, faster first drafts of related-work or evidence summaries, detection of contradicting findings, continuous monitoring of new publications.
• Evidence (with caveats): PaperQA2 (FutureHouse) is reported to match or exceed subject-matter experts on three literature-research tasks incl. the LitQA2 benchmark, and produced cited summaries rated more accurate than Wikipedia articles; flagged contradictions were validated by experts in 70% of cases [1]. The open OpenScholar-8B system beat GPT-4o on correctness by 5% on ScholarQABench, and GPT-4o fabricated citations in 78-90% of cases without retrieval [2]. Elicit advertises "up to 80% time savings" for systematic reviews (vendor claim, not independently verified) [3]. Benchmarks are mostly STEM/biomedical and English; treat them as upper bounds for your domain.

Variants

• Academic / student: related-work sections, thesis background, citation tracing (Zotero-centred).
• Systematic & scoping reviews: protocol-driven screening, dual-reviewer logic, PRISMA flow documentation.
• Pharma / biotech R&D: target and competitor landscaping, safety signals, internal + external corpora.
• Industrial R&D / SMB: technology scouting, materials/process literature, standards context (see KNOW-009).
• Clinical evidence: guideline and trial evidence for clinicians - see KNOW-010; patents - KNOW-007.
• Monitoring: weekly alerts on new papers for a topic with AI-written digests.
• Private use: health or consumer questions answered from peer-reviewed sources; always with the caveat of no medical advice.
• Deep research: general web-based market and desk research is KNOW-015; this use case is scholarly-corpus-centred.

How it works

1. research question / protocol
2. query planning (LLM expands terms, synonyms, MeSH/CPC-like vocabularies)
3. retrieval (OpenAlex / Semantic Scholar / PubMed / arXiv APIs + own library; hybrid BM25 + embeddings; citation-graph expansion)
4. screening (LLM judges title/abstract vs. criteria; reranker; human spot-check)
5. full-text parsing (PDF
6. structured text, tables, references)
7. extraction & evidence table (schema-constrained, with quote + page for every cell)
8. synthesis (cited answer, contradictions, gaps)
9. verification of every citation against the index
10. export (BibTeX/RIS, protocol/audit log)
11. human review and sign-off

• Agentic retrieval: PaperQA2-style agents iteratively search, gather evidence passages, rerank and summarise before answering [1][4]. OpenScholar adds a self-feedback loop that refines drafts and adds missing citations [2].
• Citation verification: resolve each reference by DOI/ID against Crossref/OpenAlex/Semantic Scholar; check retractions. Never accept a reference the index cannot resolve.
• Human-in-the-loop: researchers own the protocol, inclusion decisions on borderline papers, interpretation and authorship. For formal systematic reviews, AI is a second screener or extractor, not the sole one.

Data preparation

The sources are mostly external and already structured: bibliographic databases and APIs (OpenAlex, Crossref, PubMed), publisher PDFs under licence, and open-access full texts. Internal sources are the personal reference library (Zotero, EndNote), PDFs scattered over laptops and file shares, internal reports and lab notebooks.

Typical gaps: PDFs without text layer or with two-column layout that breaks extraction, duplicate and mis-named files, incomplete metadata (DOI, year), no record of which papers were screened, and licence restrictions on text mining.

Preparation steps: - Consolidate PDFs into one reference manager, remove duplicates, resolve DOIs and metadata automatically. - Run OCR and layout-aware parsing on scans (PREP-006, PREP-007). - Write the review protocol with inclusion and exclusion criteria before screening (PREP-019). - Tag internal reports with project, date and status (PREP-010). - Draw a gold set of papers screened by humans to measure the AI screening (PREP-018). - Check publisher licences for AI use.

Effort is low to moderate, around 15 to 25 percent, because the scholarly indexes come ready-made. See also DOC-014 for document parsing.

Implementation paths

Which path for whom

• SMB: Path A, because established AI research tools built on open scholarly indexes work out of the box with cited answers and need no engineering.
• Enterprise: Path B, because R&D wants to connect licensed publisher content and internal reports via APIs with audit logs and own access control.
• Developer: Path C, because OpenAlex and arXiv data plus open embedding models and a local LLM give a fully controllable pipeline, if a HW3 machine is available.
• Private: Path A, because students and individuals get the best result from ready tools with a reference manager such as Zotero.
• Public: Path B on a sovereign EU cloud, because universities and authorities must comply with hosting and procurement rules and mostly lack GPU capacity.

Path A - Buy (recommended for most users)

• AI literature tools: Elicit (screening and data extraction for reviews) [3], Consensus (cited answers, "Consensus Meter", study snapshots) [5], Undermind (deep literature search), scite (citation context: supporting/contrasting), SciSpace, ResearchRabbit/Litmaps (citation maps).
• Publisher / index vendors: Scopus AI (Elsevier), Web of Science Research Assistant (Clarivate), Dimensions (Digital Science), Springer Nature and Wiley AI features - strong where the institution already licenses the content. Review tools: Covidence, Rayyan, DistillerSR.
• EU/EEA angle: Iris.ai (Norway, EEA) for R&D search; most others are US-based - check DPA, EU processing and training-on-user-data settings. Reference managers: Zotero (open source) as the hub.
• Choose A when: individual researchers, small teams, no confidential corpus. Test on a topic you know well.

Path B - Build on cloud APIs

• Data: OpenAlex (open scholarly index, REST API) [6], Semantic Scholar Academic Graph API (about 214 million papers; recommendations; SPECTER2 embeddings; key required for stable rates) [7], Crossref, PubMed/Europe PMC, arXiv, CORE. Respect each API's terms and licences.
• Models: frontier LLMs with long context and structured output (Claude, Gemini, GPT, Mistral Large from the EU provider Mistral AI) in EU regions via Azure, AWS Bedrock, Vertex AI or Mistral La Plateforme; reasoning models for synthesis, small cheap models for screening at scale.
• Orchestration: PaperQA2 library (Apache-2.0, LiteLLM-compatible) [4], LangGraph/LlamaIndex, or n8n; Docling or GROBID for PDFs [8]; reranker + vector DB (pgvector, Qdrant).
• Choose B when: custom protocols, integration with internal documents, or audit logging are needed.

Path C - Self-host (open source)

• Stack: PaperQA2 with local models via Ollama/llama.cpp/vLLM (the project warns small models struggle with its instruction-following demands) [4]; OpenScholar code, models and 45M-paper open-access datastore as a reference [2]; Docling (MIT) for parsing [8]; open embeddings (BGE-M3, Qwen3-Embedding) and rerankers; Qdrant/pgvector/OpenSearch; Zotero as library. Open LLMs: Qwen3 family, gpt-oss, Llama, Mistral open-weight models.
• Hardware: HW3 for 30-120B models (good synthesis quality); HW2 works for screening/extraction with ~8-30B models and OpenScholar-8B-class systems. Index of the open-access corpus needs several TB of storage.
• Choose C when: unpublished data, pre-patent research, or pharma IP must not leave the premises. Licensed full texts may only be processed locally within the publisher licence.

Path D - Custom training

Fine-tune only for narrow, high-volume needs: a domain screening classifier (hundreds to thousands of labelled include/exclude decisions from past reviews), a domain embedding model, or an extraction model for a fixed schema. Reviewer decisions are the natural training data. General synthesis is better served by retrieval plus a strong base model.

State of the art (2026)

• Agentic literature agents (PaperQA2, OpenScholar, Ai2's Asta agents and AstaBench evaluation environment [9], vendor deep-research modes) outperform plain chat LLMs on grounded scientific QA and cite real papers because citations come from retrieved passages [1][2].
• Benchmarks: LitQA2 (biology literature retrieval QA), ScholarQABench (multi-domain, 2,967 expert-written queries) [1][2], AstaBench [9]. Scores on benchmarks do not transfer automatically to your niche or non-English literature.
• Techniques: hybrid retrieval, citation-graph expansion, reranking, contextual summarisation of chunks, schema-constrained extraction, self-feedback on drafts, retraction checks.
• Limits and failure modes: fabricated or mismatched citations (especially without retrieval); paywalled full text is unavailable, so abstract-only evidence is overstated; bias toward highly cited, English and open-access papers; misreading of tables, effect sizes and study populations; missing negative results; prompt injection hidden in PDFs; silent recall gaps that make a "systematic" result non-reproducible. Search logs and versioned prompts are needed for reproducibility.

Shared resources & combinations

• Same RAG stack (llm-general + embedding-text + reranker + vector-db + ocr-docparse): KNOW-002 company RAG, KNOW-001 semantic search, DOC-014 document parsing for RAG - one index/GPU server (HW3) can serve them.
• Neighbouring research assistants: KNOW-006 legal research, KNOW-007 patents, KNOW-010 clinical knowledge, KNOW-015 deep-research agent - same agent framework, different corpora and compliance rules.
• Downstream: KNOW-013 (knowledge graph/GraphRAG from extracted entities), SCI-011 (AI scientist and hypothesis agents consume the review output), EDU-001 (study support).
• Shared building blocks: Zotero/DOI resolver, Crossref/OpenAlex mirror, citation-verification service, PDF parsing pipeline, evaluation set of questions with known answers.

Effort, cost & KPIs

• Path A: hours to days; per-seat subscriptions, institutional licences possible.
• Path B: 1-3 months for a validated pipeline; cost drivers are full-text tokens, API calls, reviewer time.
• Path C: 2-4 months; GPU server, storage for the corpus, ongoing index updates.
• Path D: additional 1-3 months plus labelling.
• KPIs: recall against a hand-built gold set of known relevant papers; screening agreement with human reviewers (kappa); citation validity rate (target 100% resolvable); extraction accuracy per field; time per review stage; user-reported corrections.

People & change

Involve the head of R&D or the research group lead as process owner, the researchers or students as users, the library or knowledge management for licences and reference tools, and IT for access and hosting. Management decides budget and publisher licence questions.

Works council: normally no, since the tool supports individual research and does not evaluate staff. If the vendor logs search behaviour per employee, §87(1) no. 6 BetrVG may apply (BAG: objective suitability suffices), so ask the vendor and inform the council under §90(1) no. 3 where research processes change. Universities and authorities involve the Personalrat where relevant.

Data protection: published papers carry little personal data, but internal reports and clinical material can; involve the DPO, sign an AVV (Art. 28 GDPR) and consider a DPIA for health data (GUIDE-005).

AI literacy: support users with training on hallucinated references and a rule that every cited paper is opened and checked (Art. 4 AI Act, GUIDE-006). Start with a pilot group of a few researchers on one past review and compare results with the manual one. Typical concern: research integrity and authorship; address it with a disclosure policy for AI use and a documented search log. Humans decide on inclusion and write the conclusions (GUIDE-007).

Risks, regulation & compliance

• EU AI Act: minimal risk in typical research use (Art. 4 AI literacy applies). Obligations change if outputs feed regulated decisions (e.g. clinical evidence in a medical device, MDR) or text is published to the public without human editorial review (Art. 50 transparency for AI-generated text on matters of public interest, applicable from 2 Aug 2026; exempt with human review and editorial responsibility) [10]. Chatbot interfaces need AI disclosure.
• Copyright / TDM: the EU DSM Directive 2019/790 permits text and data mining for research organisations (Art. 3) and, subject to rightsholder opt-out, for others (Art. 4) [11]. Publisher licences may restrict uploading PDFs to third-party AI services - check contracts before Path A/B with licensed content.
• GDPR: papers mostly contain author names only; personal data arises with internal documents, clinical or survey data, and user query logs. Use DPAs and EU hosting, and disable vendor training on your inputs.
• Research integrity: AI must not be listed as author; disclose use per journal and funder (DFG, Horizon Europe) policies; humans verify every cited claim. Peer-review confidentiality: never upload manuscripts under review into public tools.
• Mitigations: verified-citation pipeline, retrieval-only answers, logged searches, human sign-off.

Getting started

1. Pick one recent review question whose answer you already know; build a gold set of 20-50 relevant papers.
2. Run it through two Path A tools and compare recall, citation validity and missed papers.
3. Connect Zotero and OpenAlex/Semantic Scholar to a small PaperQA2 prototype (Path B) for your own PDFs.
4. Add automatic DOI/retraction verification and a mandatory quote-per-claim rule.
5. Write a usage policy: allowed tools, confidential material, disclosure in publications, human sign-off.

Sources

1. Skarlinski et al., "Language agents achieve superhuman synthesis of scientific knowledge" (PaperQA2, LitQA2) - https://arxiv.org/abs/2409.13740
2. Asai et al., "OpenScholar: Synthesizing Scientific Literature with Retrieval-augmented LMs" - https://arxiv.org/abs/2411.14199
3. Elicit product page - https://elicit.com/
4. PaperQA2 repository (Future-House/paper-qa, Apache-2.0) - https://github.com/Future-House/paper-qa
5. Consensus - https://www.consensus.app/
6. OpenAlex help and API documentation - https://help.openalex.org/
7. Semantic Scholar Academic Graph API - https://www.semanticscholar.org/product/api
8. Docling project - https://github.com/docling-project/docling
9. Ai2 Asta (agents and AstaBench) - https://allenai.org/asta
10. EU AI Act Article 50 (transparency obligations) - https://artificialintelligenceact.eu/article/50/
11. Directive (EU) 2019/790 on copyright in the Digital Single Market - https://eur-lex.europa.eu/eli/dir/2019/790/oj
12. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj