UNDRESS: Unstructured Document Retrieval

• UNDRESS is a retrieval paradigm that converts implicit document features into structured indices (e.g., layout graphs, annotation schemas) for precise querying.
• It applies varied techniques—from layout segmentation in newspapers to neural indexing and multimodal fusion—demonstrating high recall and precision in diverse applications.
• The system design decouples offline indexing from online query execution, enabling cost-aware, adaptive schema induction and robust retrieval performance.

Searching arXiv for the cited UNDRESS-related papers to ground the article. UNDRESS is used in the literature as an acronym for “UNstructured Document RETRieval System” or “UNstructured Document REtrieval SyStem”. In the cited work, it denotes systems for retrieving, extracting, or querying information from documents whose structure is implicit rather than explicitly encoded: scanned newspaper pages, PDFs, web pages, enterprise text, financial filings, and multimodal corpora. Unstructured documents are described as data that “lack a predefined format” (Kurowski et al., 15 Jul 2025) or whose “lack of explicit organization hinders precise information retrieval” (Lin et al., 3 Apr 2026). The resulting design space spans layout-graph matching (Bansal et al., 2016), rule–NLP–LLM cascades for robotic process automation (Kurowski et al., 15 Jul 2025), annotation-driven semantic retrieval over induced schemas (Lin et al., 3 Apr 2026), sequence-to-sequence corpus indexing (Wang et al., 2022), multi-resolution embedding and graph reranking (Cakaloglu et al., 2019, Cui et al., 2022), multimodal hybrid retrieval with a modality-aware knowledge graph (R et al., 16 Oct 2025), and extraction-plus-Text-to-SQL systems for financial KPIs (Choi et al., 25 May 2025).

1. Concept and scope

A central distinction in this literature is where structuralization occurs. “AnnoRetrieve: Efficient Structured Retrieval for Unstructured Document Analysis” replaces the usual “retrieve-then-extract” or “extract-all-then-query” paradigms with a “structure-first, then query” approach: SchemaBoot induces a compact, query-oriented schema $\mathcal S^*$, the Annotator populates an annotation store $\mathcal A$, and Structured Semantic Retrieval (SSR) translates a natural-language query into an SQL-like plan over $\mathcal A$, with no LLM calls on the critical path (Lin et al., 3 Apr 2026). A different but related decomposition appears in the 2016 UNDRESS system, which separates an off-line indexing pass over the corpus from an on-line, sketch-based sub-layout retrieval pass (Bansal et al., 2016). In financial retrieval, the same general pattern becomes document ingestion, extraction into a centralized KPI database, and query execution through a Text-to-SQL agent (Choi et al., 25 May 2025).

These designs suggest that UNDRESS is less a single algorithm than a recurring systems principle: convert latent document regularities into an indexed representation that can be queried more precisely than raw text alone. The indexed object may be a 2D layout graph, a structured annotation store, a prefix tree of document identifiers, a concept map, or a knowledge graph. What remains constant is the attempt to trade unstructured surface form for a retrievable intermediate representation.

2. Layout-centric retrieval from document images

The earliest explicit UNDRESS formulation in the supplied corpus is a layout-based system for scanned newspaper pages (Bansal et al., 2016). Its pipeline begins with grayscale conversion and binarization by Otsu’s method, followed by ruling-line detection and removal through connected-component analysis. The page is then segmented into text versus non-text regions by multi-level morphological processing. ARLSA (Adaptive Run Length Smoothing) merges proximate text pixels into rectangular text blocks, after which a vertical “symmetry maximization” heuristic merges over-segmented blocks when alignment, average character height, proximity, and ruling-line constraints are satisfied. To tolerate segmentation error, the system generates multiple, domain-specific segmentation hypotheses, including removal of small “noise” blocks, merging of adjacent non-text regions, and dropping of caption bars.

Each hypothesis yields a planar directed graph $G=(V,E)$ whose nodes are layout blocks and whose edges encode immediate spatial adjacency in the four directions. Each block record stores its type $\tau_i \in \{\text{text}, \text{non-text}\}$, bounding-box dimensions, average character height, a spatial tag $$\sigma_i \in \{\text{top}, \text{bottom}, \text{left}, \text{right}, \text{center}\}$$, and neighbor identifiers. For pruning, each block is assigned a fixed-length “context string” $$k_i=[\tau_i;\#\text{neighbors}_\text{top};\#\text{neighbors}_\text{bottom};\#\text{neighbors}_\text{left};\#\text{neighbors}_\text{right};\sigma_i]$$, which is hashed into a bucketized index. At query time, a user sketches one or more sub-layouts, optionally marking blocks as “don’t care” or leaving gaps to permit partial matching. Retrieval then proceeds by hash-guided candidate generation and a graph-isomorphism–inspired matching routine with allowances for missing blocks.

The reported evaluation uses 4,776 newspaper pages, with ARLSA succeeding on 93.1% of an original 5,128 (Bansal et al., 2016). For single sub-layout queries, Type 1 queries yield Recall 95.4%, Precision 98.1%, and average query time 0.685 s; Type 4–6 queries with missing blocks yield Recall 91.4%, Precision 98.4%, and 0.742 s. The same study states that without multiple hypotheses recall falls below 40% while precision remains about 98%, making multi-hypothesis segmentation the decisive robustness mechanism. The paper also states clear domain dependence: its heuristics were tuned for newspaper conventions, and magazines, forms, or scientific articles may require different rules or higher-level priors.

3. Structure induction, extraction cascades, and SQL mediation

In annotation-driven UNDRESS, the structural intermediary is explicit and query-oriented. AnnoRetrieve defines candidate schemas as $\mathcal S=(\mathcal F,\mathcal H)$, where $\mathcal F$ are fields and $\mathcal H$ is a 3-tier hierarchy $\mathcal A$0, and optimizes

$\mathcal A$1

subject to constraints including $\mathcal A$2, $\mathcal A$3, $\mathcal A$4, and $\mathcal A$5 (Lin et al., 3 Apr 2026). SchemaBoot discovers multi-granularity field patterns, evaluates coverage, discriminativeness, consistency, and query matching, and uses NSGA-II to select $\mathcal A$6. SSR then parses a query into schema predicates and extract predicates, and executes

$\mathcal A$7

The crucial operational claim is that EXTRACT is invoked only on a small candidate set, while online retrieval itself uses no LLM.

A more classical enterprise-oriented variant appears in the RPA prototype UNDRESS (Kurowski et al., 15 Jul 2025). Its text-extraction module uses OpenCV preprocessing, Tesseract OCR via pytesseract, language detection through langdetect, a pyspellchecker stage, and OpenAI GPT-3.5-turbo prompted with “Correct spelling mistakes in the following text if there are any.” After tokenization and lemmatization, the pipeline reaches Jaccard $\mathcal A$8 on resumes. Information retrieval then follows a strict cascade: fuzzy regular expressions for static fields, named entity recognition for dynamic fields, and GPT-3.5 as the final fallback. The system reports spaCy precision of approximately 0.79 for nl_core_news_lg and approximately 0.90 for en_core_web_trf.

Financial UNDRESS pushes this logic toward database-centered analytics (Choi et al., 25 May 2025). The system consists of an Extraction Agent and a Text-to-SQL Agent connected through Delta Lake / Databricks SQL tables Documents, Kpis, Periods, and Qualifiers. The extraction task maps candidate spans to normalized KPI tuples $\mathcal A$9, and rule injection handles ranges, units, and fiscal-period alignment; for “22–24%,” the midpoint is computed as $\mathcal A$0 while retaining range_min=22 and range_max=24. The Text-to-SQL Agent classifies intent, injects schema metadata and sample rows, generates candidate SQL, validates syntax and semantic constraints, and ranks candidates by

$\mathcal A$1

This architecture explicitly targets analysts who need structured access to filings without direct familiarity with database schema.

4. Neural indexing, embeddings, and graph representations

A distinct branch of UNDRESS dispenses with explicit intermediate tables and instead learns retrieval directly. “A Neural Corpus Indexer for Document Retrieval” models retrieval as sequence generation: a Transformer encoder reads the query, a Prefix-Aware Weight-Adaptive decoder emits a sequence of document-ID tokens, and constrained beam search traverses a prefix tree of valid identifiers (Wang et al., 2022). The probability factorization is

$\mathcal A$2

and the decoder adapts its output weights to each prefix. Document identifiers are not arbitrary strings but “semantic IDs” obtained by hierarchical $\mathcal A$3-means with $\mathcal A$4 over BERT document embeddings. Training combines teacher-forcing cross-entropy with a consistency regularizer, and inference returns the highest-scoring valid IDs.

Another line retains dense retrieval but enriches document vectors. “A Multi-Resolution Word Embedding for Document Retrieval from Large Unstructured Knowledge Bases” constructs token representations by weighting internal layers from multiple embedding sources such as BERT, ELMo, and fastText, then aggregates them into a multi-resolution embedding (Cakaloglu et al., 2019). Passage vectors can be obtained by pooling or by ConvRR, a residual convolutional retrieval model trained with a triplet loss. The paper’s central claim is that lower, middle, and upper layers encode different context scopes and that ensembling across models captures complementary semantic views. This is an explicitly retrieval-oriented alternative to fixed single-model embeddings.

Graph-based textual UNDRESS replaces raw text with document-level concept maps (Cui et al., 2022). Documents are converted into phrase co-occurrence graphs by sentence splitting, POS tagging, constituency parsing, NP/VP extraction, lemmatization, and sliding-window edge construction. Rather than relying on structure-heavy GNNs, the study proposes semantics-oriented graph functions: N-Pool over node embeddings, E-Pool over edge endpoint pairs, and RW-Pool over random walks of length $\mathcal A$5. In a two-stage pipeline, BM25 retrieves top-$\mathcal A$6 candidates ($\mathcal A$7), and graph embeddings are re-ranked by cosine similarity to a query embedding. The paper’s empirical conclusion is notable: complex structure-oriented GNNs such as GIN and GAT underperform the simpler semantics-oriented functions on CORD-19.

5. Multimodal systems and empirical landscape

Multimodal UNDRESS generalizes retrieval beyond text-only corpora. “Multimodal RAG for Unstructured Data: Leveraging Modality-Aware Knowledge Graphs with Hybrid Retrieval” presents MAHA, which embeds text with BERT or Sentence-BERT, images with CLIP, tables via serialized text encoders, equations with an equation encoder or BERT on tokens, and charts through caption text plus GNN-derived structural features (R et al., 16 Oct 2025). A modality-aware knowledge graph stores cross-modal entities and relations using an adjacency tensor $\mathcal A$8 and node embedding matrix $\mathcal A$9. Retrieval combines FAISS-based dense search with graph traversal such as BFS or Personalized PageRank, and fuses them through

$G=(V,E)$0

This system makes explicit a theme already latent in other UNDRESS designs: retrieval is improved when semantic similarity is coupled with explicit structure.

System	Setting	Reported result
Layout UNDRESS (Bansal et al., 2016)	4,776 newspaper pages	Type 1: Recall 95.4%, Precision 98.1%, 0.685 s
AnnoRetrieve (Lin et al., 3 Apr 2026)	150 queries across 3 datasets	F1 0.87; LLM Cost 29.4 K tokens; Latency 3.2 s
NCI-Base (Wang et al., 2022)	NQ320k and TriviaQA	Recall@1 65.86; Recall@5 90.49%; R-Precision 73.90
MAHA (R et al., 16 Oct 2025)	multimodal evaluation	Recall@3 0.81; MRR 0.74; ROUGE-L 0.486; Modality Coverage 1.00
RPA UNDRESS (Kurowski et al., 15 Jul 2025)	resumes and invoices	Jaccard 0.99 / 0.81; e-mail 1.00 accuracy; total amount 0.65 accuracy
Financial KPI system (Choi et al., 25 May 2025)	1,200 SEC filings and 312 analyst queries	Extraction F1 91.2%; Query Accuracy 94.8%; Throughput Gain 2.4×

Other retrieval-oriented results further broaden the empirical picture. Multi-resolution embeddings with ConvRR reach recall@1=52.32% and recall@5=75.68% on SQuAD, and recall@1=50.67% and recall@5=67.38% on QUASAR-T (Cakaloglu et al., 2019). On CORD-19/TREC-COVID, RW-Pool reaches NDCG@20 of 51.41%, compared with 45.91% for BM25, while GIN and GAT score 29.91% and 39.49%, respectively (Cui et al., 2022). Collectively, these results show that UNDRESS performance depends strongly on the chosen surrogate structure: layout graphs, annotation schemas, semantic IDs, concept maps, and multimodal knowledge graphs each favor different operating points in recall, precision, cost, latency, and modality coverage.

6. Limitations, reporting nuances, and open directions

The literature is consistent in treating robustness as the central difficulty. The layout-based UNDRESS depends on ARLSA and symmetry maximization, and explicitly cannot recover overlapping text lines that have already merged into a single block; complex non-rectilinear graphics may also produce erroneous non-text blocks, and the segmentation rules are domain-specific to newspapers (Bansal et al., 2016). AnnoRetrieve shifts complexity offline, but it also reports initial annotation overhead and notes that the schema may need re-induction if the domain shifts (Lin et al., 3 Apr 2026). The RPA prototype observes that duplicate or misaligned labels, such as multiple “Total amount” fields, can confuse the regex–LLM pipeline, and proposes that spatial or layout analysis could further improve invoice accuracy (Kurowski et al., 15 Jul 2025). The financial KPI system identifies temporal logic, multilingual filings, human-in-the-loop review, and deeper arithmetic as remaining problems (Choi et al., 25 May 2025).

There is also a methodological tension over how much explicit structure is desirable. AnnoRetrieve argues against both coarse-grained vector matching and graph-based methods with high computational overhead (Lin et al., 3 Apr 2026), while MAHA explicitly combines dense vectors with a modality-aware knowledge graph (R et al., 16 Oct 2025). The CORD-19 study refines this tension: graph structure is useful, but semantics-oriented pooling functions outperform structure-oriented GIN and GAT in the reported setting (Cui et al., 2022). A plausible implication is that explicit structure helps most when it is tightly aligned with the retrieval objective rather than introduced for its own sake.

Reporting nuances are also visible. In the AnnoRetrieve material, the comparison section states “sub-second structured queries,” whereas the quantitative table reports 3.2 s average latency over 150 queries (Lin et al., 3 Apr 2026). Such discrepancies do not negate the broader result that online LLM usage is eliminated, but they do indicate that latency claims in UNDRESS-style systems should be read alongside the exact workload and metric definitions.

Open directions are explicitly named across the corpus. AnnoRetrieve lists dynamic or adaptive schema induction in evolving corpora, multi-modal schema induction, learned extractors that reduce annotation time, and cost-aware query planners over the annotated index (Lin et al., 3 Apr 2026). The RPA prototype proposes learned fusion-weighting for rule-versus-model scores, embedding-based retrieval, and prompt-tuned LLMs to reduce hallucinations (Kurowski et al., 15 Jul 2025). The financial system proposes cross-document aggregation and multilingual extension (Choi et al., 25 May 2025). The multimodal line already points toward a retrieval substrate in which dense similarity, graph traversal, and modality coverage are jointly optimized (R et al., 16 Oct 2025). Taken together, these directions suggest that future UNDRESS research will continue to move from ad hoc document handling toward explicit, queryable, and cost-aware representations of formerly unstructured data.