Improving Access to Historical Archives with Real-time RAG-based Systems

Finding Change in Satellite Archives from Text: How to Combine Before-and-After Images Efficiently

Operational Earth observation increasingly calls for answering queries such as ``find the image pairs where a new building appeared.''This means searching an archive of before-and-after (bi-temporal) satellite image pairs and ranking each pair by how well it matches a natural-language description of the change. The component that performs this match, the fusion module that combines the ``before''and ``after''views, must be run at query time across many candidate pairs, so its speed largely sets the cost of every search. We present a controlled comparison of how to build that module. Using one fixed image encoder (a frozen CLIP model) and one training recipe for all variants, we evaluate eight designs drawn from three families: attention, state-space models (Mamba), and learned compression (our Temporal Bottleneck Fusion, TBF). Each design is tested on two benchmarks (LEVIR-CC and Dubai-CC) with ten random seeds, so the reported differences are statistically grounded. We outline three findings: first, a training-free two-stage search (a cheap difference model that shortlists candidates, followed by attention fusion that re-ranks them) matches or exceeds full-fusion recall on LEVIR-CC while cutting query cost $10$-$15\times$, with comparable R@1/R@5 on Dubai-CC; second, the linear-time scan of Mamba, attractive on paper, gives no speed benefit at the patch counts typical of vision transformers ($L{=}196$): the scan is limited by memory bandwidth, whereas attention maps cleanly onto parallel hardware; and third, compressing the fused representation (TBF) reduces parameters by $2.3\times$ and latency by $1.6\times$ for a change-only BLEU-1 cost of $0.007$, although more aggressive compression quietly discards change-relevant detail that aggregate metrics fail to reveal.

A Multimodal Retrieval-Augmented Generation Framework for Context-Aware Semantic Querying

Retrieval-Augmented Generation (RAG) has established itself as a compelling strategy for grounding large language model outputs in documentary evidence. However, production deployments continue to rely almost exclusively on homogeneous text corpora, even as enterprise repositories grow increasingly heterogeneous blending technical schematics, radiological images, annotated diagrams, and unstructured prose within the same archival system. This mismatch between system design and data reality motivates the present work. We propose a Multimodal RAG framework that unifies text and image retrieval through four tightly coupled components: a dual-stream embedding engine, a learned four-class AI query router, an adaptive confidence threshold, and a session-aware context store. Text is encoded with the allmpnet-base-v2 Sentence Transformer; images are embedded in the same 768-dimensional space via a domain-adapted CLIP ViT-L/14 model fine-tuned on approximately 120,000 technical and clinical text-image pairs. Concatenating 768-d vectors yields a 1,536-d composite query that drives a single approximate nearest-neighbor (ANN) search simultaneously across both modalities. A fine-tuned DistilBERT router assigns each query to one of four retrieval pathways text-only, image-only, hybrid, or conversational at 91.3% accuracy. Experiments were run on three enterprise corpora (engineering manuals, clinical case summaries, and legal paperwork) with the results being a mean F1@5 of 0.90, an increase of 22% over a dense text-only baseline, and a reduction of 49% in the number of hallucinations. The median first-token latency is 1.34s, which meets the interactive-use target deployment-contexts. These results show that it is possible to implement modality-aware retrieval in a simple and realistic setting with real-world organizational constraints and that the results are reproducible.

ADAPTIVE MULTI-STAGE VECTOR RETRIEVAL FOR RETRIEVAL-AUGMENTED GENERATION

Retrieval-Augmented Generation (RAG) is the standard for grounding Large Language Models (LLMs), but its parts' embeddings, lexical scoring, query expansion, reranking, and adaptive sizing do not always compose additively. Under default configurations, adding naïve hybrid retrieval with Reciprocal Rank Fusion degrades NDCG@10 by up to 5.6% on scientific corpora; temperature-sampled query expansion with a small LLM further degrades it by 16.8%. To address this, we propose the Adaptive Multi-Stage Vector Retrieval (AMSVR) framework, prioritising weighted, drift-resistant composition over uniform fusion. Rather than a one-sizefits-all system, AMSVR offers tailored configurations: AMSVR-Scientific (dense + tuned hybrid) peaks at NDCG@10 = 0.7570 on SciFact, while AMSVR-Full (seven stages) targets broader, noisier corpora where Recall@100 matters most. We evaluate on three BEIR benchmarks (SciFact, NFCorpus, FIQA), release an offline configuration-diagnostic tool, and provide a per-corpus recipe for selecting which stages to enable.

Large Language Models as Digital Libraries: A Multi-Benchmark and Multi-Model Study

Querying LLMs as digital libraries is feasible, but its effectiveness depends on model strength, deployment conditions, dataset structure, and execution strategy, and Galois remains valuable when relational discipline and controlled query execution are required.

BANANA: Bounded Adaptive Navigation Architecture for Nested Archives

Retrieval-augmented question answering (QA) over enterprise PDF archives frequently produces confident near-miss answers because dense embeddings discard entity identity, numerical units, and document provenance. Plain-text indices preserve the entity, unit, and provenance signals that dense embeddings discard. Banana (Bounded Adaptive Navigation Architecture for Nested Archives) replaces opaque vectors with plain-text Markdown indices built via one-time VLM page transcription, eliminating both a separate Optical Character Recognition (OCR) engine and a vector store. At query time it applies Progressive Retrieval with Overlap (PRO), an iterative selector with a deterministic contraction bound, followed by a K-adaptive extraction gate that escalates retrieval depth only upon evidence insufficiency. Across three benchmarks, Banana consistently outperforms the strongest baseline: 71.4~EM on TAT-DQA (+15.2 over PageIndex; p <0.001), within 5.2 of oracle TAT-LLM; 79.3% accuracy on FinanceBench-150 across 53K~pages (+17.3 over PageIndex; p <0.01); and 86.2% Completeness@10 on UniDoc-Bench across 8~domains (+20.8 over the strongest fusion baseline). All gains hold across multiple backbones, including a local model at zero API cost.

CroQS: Cross-modal Query Suggestion for Text-to-Image Retrieval in Image Archives