Research
DeepView is a Retrieval-Augmented Generation (RAG) system. Each of its components implements a method published and evaluated by the research community: lexical and dense retrieval, rank fusion, re-ranking, knowledge graph extraction, verification of answer groundedness.
We only draw on work from the field's leading conferences and teams, and only where it holds up against real documents. Each publication has its own entry: the problem it formalises, the method it proposes, and what our implementation retains from it, including the departures we make from the original work.
Cui et al., “RT-DocLayout: Real-Time End-to-End Document Layout Analysis with Reading Order in the Wild”, 2026 · arXiv:2606.23344
A corporate document is rarely a clean page: it is scanned crooked, photographed, creased. Multi-model pipelines accumulate errors from one stage to the next.
In DeepView: layout analysis on degraded documents
Read the article →Śmigielski et al., “Chunking Methods on Retrieval-Augmented Generation: Effectiveness Evaluation Against Computational Cost and Limitations”, Wrocław University of Science and Technology, 2026 · arXiv:2606.00881
Splitting documents into passages looks like a technical setting with no consequences. It is in fact one of the choices that weighs most on what a system can retrieve.
In DeepView: the choice of chunking strategy
Read the article →Auer et al., “Docling Technical Report”, IBM Research, 2024 · arXiv:2408.09869
PDF encodes a rendering, not a structure: glyphs positioned by coordinates. Reading order, columns and tables must be reconstructed by inference.
In DeepView: layout analysis at indexing time
Read the article →Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, NeurIPS 2020 · arXiv:2005.11401
A language model's knowledge is frozen in its parameters at training time. Queried outside that distribution, it generates a plausible answer rather than a refusal.
In DeepView: the overall system architecture
Read the article →Robertson & Zaragoza, “The Probabilistic Relevance Framework: BM25 and Beyond”, Foundations and Trends in Information Retrieval, 2009
Dense representations project rare terms into a continuous space where an identifier loses the specificity that made it discriminating. Lexical weighting remains necessary.
In DeepView: the lexical component of retrieval
Read the article →Reimers & Gurevych, “Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks”, EMNLP 2019 · arXiv:1908.10084
Lexical matching fails as soon as two phrasings of the same concept share no terms. The problem is vocabulary mismatch between query and document.
In DeepView: the dense component of retrieval
Read the article →Nogueira & Cho, “Passage Re-ranking with BERT”, 2019 · arXiv:1901.04085
Independent encoders are indexable but imprecise: query and passage are vectorised separately, with no interaction between their terms.
In DeepView: re-ranking of candidates
Read the article →Cormack, Clarke & Büttcher, “Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods”, SIGIR 2009
A BM25 score and a cosine similarity share neither scale nor distribution. Combining them linearly assumes a normalisation that nothing justifies.
In DeepView: rank fusion
Read the article →Edge et al., “From Local to Global: A Graph RAG Approach to Query-Focused Summarization”, Microsoft Research, 2024 · arXiv:2404.16130
Passage retrieval treats each unit independently. A query whose answer is distributed across dozens of documents exceeds what that paradigm can assemble.
In DeepView: graph extraction and traversal
Read the article →Asai et al., “Self-RAG: Learning to Retrieve, Generate and Critique through Self-Reflection”, ICLR 2024 · arXiv:2310.11511
A relevant context does not constrain generation. The model can produce an assertion not entailed by the supplied passages, with nothing to flag it.
In DeepView: groundedness and relevance verification
Read the article →Yan et al., “Corrective Retrieval Augmented Generation”, 2024 · arXiv:2401.15884
Generation quality is bounded by retrieval quality. A system that generates unconditionally converts retrieval failures into fluent, false answers.
In DeepView: candidate assessment before generation
Read the article →This work belongs to its authors and has no connection to Alister AI: we claim neither authorship nor endorsement, and none of these authors contributed to DeepView. We select a publication on two criteria: how solid its evaluation is, and whether it holds up on real documents rather than on academic benchmarks alone. A gain that does not survive contact with corporate documentation is of no interest to us, however prestigious the venue. That is also why we sometimes depart from the original method where our deployment constraints call for different choices: those departures are flagged in the entries concerned.
Evaluation
We evaluate what the knowledge graph adds over a purely vector RAG architecture, on a question set stratified by complexity. The protocol is described below; the campaign is under way.
the answer appears verbatim in a single passage.
the question and the relevant passage share no vocabulary.
the answer requires combining several passages or documents.
the answer requires following a chain of relations between entities.
questions built so that vector similarity alone is insufficient.
The evaluation campaign is under way. We will publish the full study here: detailed protocol, question set, and per-level, per-metric results. Until then, the chart opposite represents our working hypothesis, not measurements.
Architectural choices, the departures we make from the literature, and interim evaluation results are discussed directly with the team that builds the system.