Domain-specific RAG & cross-corpus question answering
210 evidence items
AI that performs retrieval-augmented generation over proprietary domain-specific corpora and answers questions across multiple knowledge bases. Includes specialised embedding and retrieval for technical domains; distinct from enterprise search which targets general internal documentation.
Overview
Domain-specific RAG grounds a model's answers in proprietary, specialised corpora and reasons across several knowledge bases at once, using retrieval tuned to the vocabulary and structure of a technical field. It matters to anyone whose valuable knowledge sits outside general documentation, and it is a leading-edge practice and steady: production tooling is generally available and named deployments report real gains, yet no repeatable adoption path has emerged. Every success rests on bespoke per-domain engineering, including curated sources, adapted retrieval and content governance, while failures often go undetected as corpora degrade or grow. Until that tuning burden shrinks and independent analysts endorse the practice as mature, it stays an expert's discipline.
Current Landscape
Production deployments span healthcare, regulated energy, telecommunications, legal contracts, scientific publishing and particle physics. At CERN, the deployed chATLAS assistant gives conversational access to more than 30,000 internal ATLAS documents through hybrid dense-sparse retrieval. Its authors report that domain fine-tuning combined with contextual retrieval more than doubles top-1 correct-document retrieval on the hardest source. A health-education conversational agent for gastric cancer patients reached 82% accuracy and a 100% RAG knowledge hit rate after iterative alpha testing.
Regulated and commercial deployments report their gains with less supporting detail. Ontario Power Generation formalised its progressive evidence acquisition pattern as PEA-CAE for energy-board filings. A telecommunications support deployment reports 71% ticket deflection. Advancing Analytics claims 96% faster, 99% cheaper and 90% accurate contract extraction on Databricks, with no methodology given. None of these figures has been independently replicated.
In the clinical case, retrieval rather than agentic orchestration drives the measured gains. Dr.Copilot's developers report a score of 4.557 against 4.217 for a fixed backbone on a 75-case external validation, with clinically important errors at 5/75 against 16/75. In their ablation, standard RAG scored highest at 4.813, and added tools and model-directed routing gave no uniform gains. The authors say agentic orchestration does not itself establish clinical benefit or patient-safety improvement.
Graph and entity structures laid over a corpus now outperform flat retrieval on cross-document questions. Graph-RAG beats vector-only retrieval on multi-document reasoning and cross-document trustworthiness. The CorpusMap authors report quality gains of 6.4 to 11.7 points with 34% to 57% fewer input tokens than raw-corpus agentic search, across 7 models and 3 benchmarks. The PILAR authors report +1.6 EM over flat retrieval, rising to +5.9 on 3-hop questions. Knowledge-graph RAG, evaluated on 661 query-context-answer triplets from one building code, outperformed pretrained transformers and general-purpose LLMs in answer quality.
Graph structure can buy recall at the expense of faithfulness. ClinLab-RAG's developers, working with a clinical laboratory distributor in Ecuador, report answer relevancy rising from 0.416 to 0.522 over keyword retrieval while faithfulness fell from 0.945 to 0.724. Their revised version checks compatibility deterministically against the manufacturer's matrix. Complete GHS classification then appeared in 14 of 17 safety answers, against 1 of 17 before. Three blinded domain experts rated it near ceiling, not significantly above the strongest baseline.
Domain-specific reranking improves evidence selection more than it improves answers. A study of question answering over longitudinal clinical notes from 200 bariatric surgery patients found that MedCPT cross-encoder reranking raised Hit@10 from 46.6% to 60.6%. Judge-assessed answer correctness rose only from 44.8% to 48.6%. The authors caution that gains in retrieval do not translate proportionally into gains in answer correctness. Extraction in technical domains stays below 40% exact-match accuracy.
Domain-specialised encoders are not a safe default. A Scientific Reports study of question answering over a single medical textbook found general-purpose embeddings, particularly Jina-Embeddings-v3 and BGE-base-en-v1.5, outperformed biomedical-specific encoders in both retrieval and generation. Embedding choice mattered more than chunking strategy. Grounding stayed incomplete: 35.6% of answered questions lacked explicit evidence spans, and the best configuration still left 26.1%.
Embedding choice can also outweigh model choice. An independent evaluation of 22 open-weight models on a Croatian Chart of Accounts benchmark found that a multilingual encoder raised average Exact Match by about 43 points on the Croatian corpus. Retrieval cut invented account codes from 33.6% to below 0.3% under every condition, though residual errors were mostly wrong-but-existing accounts. Non-generative baselines beat every model on the English corpus and 10 of 22 on the Croatian one.
Curated corpora beat raw ones when practitioners do the scoring. The KUPAS MASTER authors report practitioner scores of 89.58 for agents using its curated experience assets, against 79.75 for raw-corpus RAG and 70.63 for the base model, over 177 questions from 20 practitioners. The figures are platform-reported and unreplicated. A Springer survey of RAG chatbots reaches the same conclusion from the literature, finding systems most effective when knowledge sources are well curated and retrieval is tailored to domain constraints.
Grounding on a curated regulatory corpus lifts faithfulness sharply in compliance prototypes. A working-at-heights construction safety system built on RAGFlow reports faithfulness of 0.45 for the pure model, 0.91 for local RAG and 0.97 with online supplementation, over 100 test questions. Latency rose from 2.1 s to 6.8 s across the same configurations. The authors built both the system and its benchmark, and report no field trial.
Corpus quality is the sharpest failure point. The RAGStress benchmark reports that a clean knowledge base lifts an 8B-parameter model from 42.3% to 98.6% accuracy. Factual corruption in the same documents drops that model to 27%, below its no-retrieval baseline. Its authors conclude that clean retrieval masks robustness differences between models. The Springer survey names irrelevant or contradictory retrieval, prompt injection through retrieved text, citation laundering and over-reliance on top-k passages as recurring failure modes.
Benchmark scores overstate accuracy on real documents. Salesforce documents a benchmark-to-production accuracy gap in its enterprise RAG pipeline, with systems exceeding 90% on standard benchmarks dropping to ~46% on real enterprise documents. The causes given are silent parsing failures, superficially similar wrong answers ranking high by semantic similarity, and header/table separation at page breaks. Structured-data handling, where queries mix database records with unstructured text, remains architecturally distinct and underserved.
Lexical retrieval often holds its own against denser stacks. A controlled study of a local RAG assistant over 18 industrial safety documents found raw BM25 had the highest mean reciprocal rank, at 0.492, and no significant difference from hybrid fusion. Its authors add that document-level evaluation can overstate access to exact evidence. LegalAssist's authors, evaluating on Indian prior-case retrieval with gold relevance judgments, found BM25 the strongest configuration. Dense retrieval, multi-query expansion and cross-encoder reranking failed to improve on it consistently.
Multilingual and culturally grounded retrieval is a distinct weak spot. Evidence shows embeddings degrade 10-40% on underrepresented languages. Generic retrieval also fails on culture-specific facts. Multilingual boundaries degrade performance by 10-23 percentage points. Current research responds with culturally aligned adaptive retrieval loops and with knowledge-graph augmentation as an alternative to retrieval for culturally grounded question answering.
Grounding helps low-resource clinical question answering without making it safe. The BanglaDiaHallu preprint reports that hybrid retrieval over Bangladesh's national diabetes guideline cut Qwen3.5-9B's hallucination rate from 77.5% to 11.3%. Its authors found no configuration free of hallucination, said base model choice mattered as much as retrieval, and described residual errors as specific, confident and potentially harmful.
Traceable citations do not make clinical answers reliable. In a local RAG study over 14,637 PubMed records on perioperative inflammatory bowel disease, all 236 retrieved PMIDs existed in PubMed. RAG significantly improved Qwen3-VL-8B's accuracy but not DeepSeek-R1-Distill-Llama-8B's, and patient safety showed no reliable benefit. Faithfulness fell with retrieval for Qwen3. The authors conclude that authentic references can accompany incorrect or unsafe conclusions, so traceability and clinical reliability need separate assessment.
Broader adoption remains blocked by per-domain engineering burden and by brittleness as corpora evolve. Production success depends on source-authority governance, corpus curation, domain-appropriate embeddings, hybrid retrieval combining BM25, dense search and cross-encoder reranking, per-domain tuning and rigorous evidence-level evaluation. None of these transfers cleanly from one domain to the next. The newest results reinforce the point: gains and losses alike trace to the state of the knowledge base, the embedding choice and the evaluation design rather than to the model or the orchestration layer.
Tier History
Evidence (210)
— Independent 22-model benchmark: a multilingual encoder adds about 43 points of Exact Match, retrieval cuts invented codes from 33.6% to below 0.3%, and non-generative baselines often win.
— Construction-safety RAG prototype on RAGFlow: faithfulness 0.45 pure model, 0.91 local RAG, 0.97 with online supplementation over 100 questions, at rising latency; self-built benchmark.
— Legal RAG on IL-PCR gold labels where BM25 beat dense retrieval, multi-query expansion and cross-encoder reranking; value lies in temporal checks, citation verification and abstention.
— Single-site medical-equipment RAG where a compatibility graph raised recall but cut faithfulness from 0.945 to 0.724; deterministic checks fixed GHS answers, yet experts saw no significant gain.
— Deployed domain RAG at CERN over 30,000+ internal ATLAS documents with hybrid dense-sparse retrieval; domain fine-tuning plus contextual retrieval more than doubles top-1 retrieval on the hardest source.
205 more · latest 2026-10-19 →
— Controlled industrial-safety retrieval study: document-level evaluation overstates access to exact evidence, and hybrid fusion does not significantly beat BM25.
— Negative and mixed: guideline-grounded RAG cut Bangla diabetes hallucinations from 77.5% to 11.3%, but no configuration was free of hallucination and base model choice mattered as much as retrieval.
— Peer-reviewed knowledge-graph RAG over a building code (661 triplets) beat pretrained transformers and general LLMs. It adds a regulatory domain to the graph-over-flat evidence.
— Consultancy talk on a contract extraction and chat system over legal documents using Databricks Vector Search. It claims 96% faster, 99% cheaper and 90% accurate extraction, all unverified.
— Negative signal on corpus quality: across 52,500 evaluations an 8B model goes from 98.6% on a clean knowledge base to 27% under factual corruption, below its 42.3% no-retrieval baseline.
— MedCPT reranking over clinical notes from 200 patients lifts Hit@10 from 46.6% to 60.6%, but answer correctness only from 44.8% to 48.6%: retrieval gains do not carry through to answers.
— An offline entity-document graph over multi-source corpora improves quality by 6.4 to 11.7 points with 34% to 57% fewer input tokens than raw-corpus agentic search, across 7 models and 3 benchmarks.
— Self-reported platform study across 20 practitioners and 177 questions: curated experience corpora score 89.58 against 79.75 for raw-corpus RAG and 70.63 for the base model.
— Peer-reviewed survey finding RAG works best with curated sources and domain-tailored retrieval, and cataloguing failure modes such as contradictory retrieval and citation laundering, plus weak evaluation practice.
— Negative and contrarian: over a medical textbook, general embeddings beat biomedical encoders, and even the best setup left 26.1% of answered questions without evidence spans.
— An entity-linked assertion graph over text, tables and figures gains +1.6 EM over flat retrieval and +5.9 on 3-hop questions; ablations show visual assertions add little without filtering.
— Benchmark research comparing Graph-RAG against standard RAG on culturally specific QA (LatamQA dataset), showing 72-78% error reduction versus base LLM and zero-shot multilingual transfer to Portuguese.
— Peer-reviewed synthesis of 104 sources documenting RAG grounding failures (citation-to-claim entailment gaps, suppressed uncertainty, causal language creep) and proposing accountability evaluation framework for evidence-grounded systems.
— Peer-reviewed ACL 2026 Findings paper on adaptive multilingual RAG with evidence-quality feedback and corpus reselection, achieving up to 3.58 percentage point accuracy improvement on low-resource languages.
— Peer-reviewed healthcare deployment across 44 patients and clinical staff showing RAG knowledge hit rates improving from 86% to 100% across test rounds and answer accuracy reaching 82% through iterative refinement.
— Self-reported production deployment formalising RAG evolution through four architectural stages into the PEA-CAE pattern (Progressive Evidence Acquisition with Cost-Aware Escalation) for regulatory compliance QA over evolving corpora.
— Self-reported production telecom network operations deployment achieving Precision@10 0.69, Recall@10 0.79, and 89.6% grounding rate through hybrid retrieval (BM25+dense+KG); +15-18 percentage point gains over vector-only baseline.
— Research on directory-aware semantic storage for structured documents achieving state-of-the-art RAG accuracy while consuming only 11.6-51.9% of tokens required by competing systems, addressing efficiency on enterprise corpora.
— Research applying Qwen2-VL-2B-Instruct multimodal extraction to generate textual summaries of scientific tables and figures, reporting 157% retrieval-quality improvement over naive RAG baseline with only 50ms additional latency.
— Production case study showing RAG systems exceeding 90% on benchmarks dropping to ~46% on real enterprise documents; remediation via intelligent parsing and reranking raised performance to 84.4%, documenting persistent benchmark-production gaps.
— Research demonstrating exact-span extraction accuracy below 40% in technical domains with current semi-extractive RAG methods; proposes constrained hybrid decoding (CHyD) to enforce hard decoding constraints on quoted spans.
— Domain-specific comparison for evidence-sensitive industries (pharma, biotech, legal): traceability and citation discipline are domain requirements distinct from accuracy; FDA/biopharma deployments mandate human verification and source provenance.
— UCSF production deployment: agentic RAG on 7M+ subject EHR with curated biomedical knowledge graph improved accuracy on 100 clinical tasks for GPT-5.5 and Claude Opus, accelerating research protocol timelines from months to hours.
— Technical guide: fine-tuning embeddings on domain data improves retrieval precision 20-40% on specialized corpora; demonstrates generic embeddings fail on domain terminology (clinical, legal, financial) requiring specialized semantic spaces.
— FinRAG-QA benchmark: 999 practitioner-curated financial Q&A on 209 regulatory documents; domain-specific optimization (retrieval-adapted embeddings, reranking) achieves +38.8% NDCG and +34.4pp generation accuracy.
— Production failure mode: corpus contradictions retrieved cleanly but outdated source ranked higher by semantic similarity, yielding confidently-cited wrong answers; solution requires source-authority governance, demonstrating domain-specific RAG success depends on content governance, not retrieval tuning.
— Industry analysis of RAG maturity ceiling: 72% of enterprise implementations fail first year; Gartner forecasts 40% agentic AI projects canceled by 2027 due to cost, unclear ROI, governance gaps; signals architectural evolution toward context-layer governance and knowledge-graphs.
— Peer-reviewed benchmark comparing vector-hybrid and graph RAG on government project management data; graph RAG superior for trustworthiness (faithfulness +3%, context recall +4%, answer relevancy +5%).
— Epoch8 production deployment: hybrid structured-query + vector retrieval for structured domains achieves 90%+ accuracy vs 40-60% for vector-only RAG; demonstrates architectural brittleness, not fixable via retrieval tuning.
— Production financial RAG: systematic 6-stage debugging (context reordering, lost-in-middle fix, claim verification, deterministic math) reduced hallucination from 15% to 1%, demonstrating pipeline discipline achieves reliability for regulated domains.
— Peer-reviewed case studies: four corporate domain-specific RAG deployments (assembly, maintenance, commissioning) identify critical success factors (knowledge base strategy, metadata enrichment, employee participation) and persistent challenges (hallucination, data inconsistency).
— Preregistered empirical study documenting production RAG fragility: corpus expansion causes silent answer churn (6.44pp exact-match excess, 10.25pp semantic excess on NQ) despite fixed model/prompt/retrieval. Compatibility failures can mask accuracy metrics—critical signal for deployment discipline.
— Domain-specific RAG system for industrial QA (manufacturing, maintenance, fault diagnosis) with multi-module pipeline: domain classifier, GTE-DPR dense retrieval, BGE reranker; demonstrates domain-adapted components essential for specialized technical corpora.
— Mistral Agentic Search GA replaces one-shot RAG with multi-step retrieval loop (search, open, navigate, read, grep); FinanceBench: +59.3pp (26.7%→86%), OfficeQA: +45.6pp (6.3%→51.9%); 39.6% latency reduction, up to 1/3 token savings.
— SmartReco domain-specific RAG for personalized recommendations: hybrid retrieval (BM25+dense), cross-encoder reranking, query rewriting; 96.3% Precision@3, 0.9739 NDCG@3 (21.16% vs semantic-only); demonstrates retrieval stratification effectiveness at production scale.
— Domain-specific RAG architecture for regulated domain (insurance) addressing six failure modes: scoped supersession, contradictions, terminology drift, temporal misalignment, rationale loss, multi-hop brittleness; Azure implementation with Cosmos DB knowledge layer and entity resolution patterns.
— Peer-reviewed evaluation of domain-specific RAG across five architectures on Bengali agricultural advisory (1,000 queries, 2,882 KB items); BM25 R@10=0.506, hybrid RRF R@10=0.539; dense retrieval collapses 10× by query type, establishing domain-specific strategy necessity.
— Accenture Japan IT operations RAG deployment: ticket triage + procedure synthesis via hybrid search (BM25+vector), LLM metadata attachment, Azure AI Search indexing; outcomes: faster incident response, new engineer autonomy, eliminated document search overhead.
— Domain-specific RAG deployment processing 1,000+ scientific documents (journals, patents, lab reports) for materials R&D via vector embeddings and semantic segmentation; zero-data-retention security with measurable time savings in literature review bottleneck.
— Named domain-specific agentic RAG deployments (AIG: SLA reduced 5x+, accuracy 75%→90%; Allianz: autonomous workflows on motor/health claims; Travelers: 10K engineers; Verisk: domain-specific data access); signals production maturity in regulated insurance sector.
— Azure Databricks ai_search() SQL function in Beta enables batch domain-specific RAG pipelines with multi-source retrieval, deduplication, reranking, and grounded answer synthesis—signals data-platform convergence of RAG infrastructure at enterprise scale.
— Analysis of production RAG failures: 73% of 143 enterprise deployments experienced critical failure within first quarter; 41% undetected by standard evals; documents cross-document entity resolution, temporal drift, and multi-hop gaps as blocking adoption in domain-specific systems.
— Peer-reviewed study of domain-specific RAG on CORD-19 scientific corpus; hybrid retrieval (BM25 + BGE-M3) achieves Recall@10 1.0, domain-mismatched rerankers reduce precision—direct evidence that domain-adapted components are essential in specialized corpora.
— Empirical study of 12,000 multi-hop QA trajectories identifies procedural failure in agentic RAG: agents retrieve evidence but skip reading before answering; Read-Gate invariant improves accuracy 14.9–19.9 points, decomposes pre- vs post-evidence failure modes.
— Xiaohongshu production deployment: domain-specific RAG for 5,300+ table data warehouse; knowledge-graph-guided retrieval improved Hit@10 from 19.1% to 96.6%, reduces tokens 71.6×, demonstrates enterprise retrieval solving real asset discovery failures.
— Controlled study spanning 1.7M to 600M tokens (28 tiers) comparing BM25, dense, graph, and agentic RAG; BM25 overtakes agents around 10M tokens, maintains 20-point margin at 600M; establishes lexical retrieval as scalable default for enterprise domain-specific corpora.
— ACL 2026 Industry Track framework addressing multi-hop QA as practical enterprise bottleneck; achieves 23.5% accuracy improvement and 10.5% NDCG gain over baselines without retraining, applicable to domain-specific cross-corpus synthesis.
— Practitioner analysis of silent production failures from embedding model swaps; proposes dual-index migration pattern for safe upgrades without coordinate system mismatch; essential governance practice for domain-specific RAG systems under continuous evolution.
— MIT/Microsoft research synthesizes CoRe-RAG framework reducing multi-hop query failures from 71% to 23% through specialized agent roles with closed-loop verification; demonstrates architectural solution to domain-specific cross-corpus reasoning bottleneck.
— Named company (Cortex) deployed domain-specific RAG on 60+ technical sources (docs, GitHub, APIs, tickets); answered 18,000+ queries with 86% certainty and 48,500+ citations, demonstrating production-scale technical domain deployment.
— Microsoft Research benchmark on 9,846 QA pairs across heterogeneous domain-specific data lakes identifies critical failure modes in cross-corpus reasoning; baseline RAG and agentic systems fail on relation chaining, policy grounding, and tabular QA.
— Production multi-agent RAG framework for SEC filings with corpus-aligned retrieval and online A/B testing on 1,000 real users; demonstrates deployment-ready domain-specific multi-agent orchestration for financial document QA.
— Security research identifies vulnerability in agentic multi-hop RAG; truth-preserving edits redirect reasoning toward wrong answers with 83.3% success rate; tested on 5 frontier models and 3 agent architectures, exposing hidden failure mode in production systems.
— UC Berkeley SkyLab and Google DeepMind MultiHopRAG benchmark evaluates 23 enterprise configurations on 10,000 multi-hop questions; only 8% achieve correct answers; 63% of failures from incomplete retrieval chains, 41% from hallucinated bridges.
— Analysis of enterprise RAG operational degradation: embedding drift, access control failure, hallucination from stale context; identifies silent failures with no error signals until user impact observed.
— Critical analysis identifying shift from vector-DB-first RAG toward agentic search with permission-aware retrieval and tool-driven orchestration; establishes new production pattern for multi-step domain-specific reasoning.
— Microsoft Foundry IQ achieved GA as unified knowledge platform for multi-agent access; serverless agentic retrieval delivers 46–54% evidence recall improvement and 34% token cost reduction.
— Production survey documenting 8-sector RAG deployments with domain-specific architectures: banking >95% citation precision, insurance 70%+ touchless claims processing, healthcare >95% faithfulness.
— Production deployment guide establishing agentic RAG as standard enterprise pattern for multi-source synthesis in finance, legal, support; quantifies 3–10× token cost increase vs traditional RAG.
— Peer-reviewed taxonomy of 33 RAG failure modes across 7 pipeline stages; identifies critical gaps—12 modes lack empirical evidence including all 8 agentic orchestration modes—revealing research frontier.
— Multi-agent RAG deployed at Brazilian Federal Police Forensic Institute for cross-corpus evidence reasoning; demonstrates domain-specific maturity in regulated, high-stakes environment with traceability requirements.
— Documents four structural RAG failure modes (similarity ≠ correctness, synthesis hallucination, meaning re-derived at query time, temporal unawareness) affecting domain-specific systems; provides balanced negative signal on architectural limitations blocking broader adoption.
— Amazon Science benchmark with 35K+ human-curated grounding annotations across 2366 questions enables fine-grained end-to-end RAG evaluation; signals Tier-1 vendor investment in production-grade domain-specific RAG benchmarking.
— MLflow RAG Agents framework reduces hallucination from 60% to <7% on enterprise clusters via five patterns: query decomposition, self-reflection, context chaining DAGs, tool-augmented reasoning, and evidence validation; demonstrates production-scale cross-corpus reliability.
— U.S. Navy warfare center production RAG deployment on AWS GovCloud demonstrates operationally mature domain-specific RAG with security controls (NIST SP 800-53, DoD RMF, CMMC) for classified data handling; validates regulated-domain viability.
— Gemini Enterprise Agentic RAG with Sufficient Context Agent (93% accuracy determining answer sufficiency) achieves 34% factuality improvement over standard RAG and 90.1% accuracy on cross-corpus scenarios; production GA feature.
— Ablation study on agentic RAG multi-hop reasoning shows fixed hybrid retrieval outperforms adaptive routing by +1.8 EM; two iterations capture 95% of gains, validating architectural simplification for cross-corpus reasoning.
— Databricks AI Search now includes native retrieval quality evaluation generating synthetic queries and scoring via LLM-as-judge on graded relevance scale (DCG@10 primary metric); maturity signal that systematic evaluation moved to production infrastructure.
— Production domain-specific RAG at billion-scale (1.3B+ profiles) using MUSE custom domain embeddings; achieves +4% HRR, −5% false positive rate, +24% Kappa quality improvement—evidence of leading-edge deployment maturity.
— Cross-query consistency framework addressing cross-corpus robustness: demonstrates semantically equivalent queries retrieve different results in multi-document corpora; solution achieves +4.76 pp EM on TriviaQA, +9.12 pp on MuSiQue.
— Multi-turn RAG evaluated across four distinct domains (finance, cloud docs, government, Wikipedia) with significant cross-domain performance variation; demonstrates domain-specific retrieval strategy selection is essential rather than universal.
— Peer-reviewed research identifying vector search dilution as critical failure mode in large cross-corpus RAG systems; proposes domain-scoped retrieval solution validated across 5 LLM backbones, 6 corpora, and named Wyoming DOTD deployment.
— Landmark analysis identifying three architectural pathologies (mereological, diachronic, causal blindness) showing RAG failures in legal domain are structural mismatches, not confabulation—directly applicable to regulated domains requiring hierarchy and temporal dynamics.
— Google Agentic RAG now in public preview on Gemini Enterprise Agent Platform with iterative multi-agent workflow; 34% factuality accuracy improvement vs standard RAG, FramesQA benchmark 90.1% on multi-corpus scenarios.
— Critical assessment of enterprise RAG failures: identifies failure occurs in retrieval layer (not generation); documents Enterprise RAG Gold Standard benchmark for realistic failure modes (multi-hop dependencies, conflicting regional policies, structured data in PDFs).
— Empirical comparison across domain-specific RAG (DevOps KB) and multi-hop reasoning benchmarks showing architecture effectiveness depends on domain characteristics; validates necessity of adaptive retrieval strategies rather than universal approaches.
— Production domain-specific RAG deployment (health insurance): 40% reduction in document lookups, 91% attribution accuracy, 6% claims accuracy improvement, handling time 14→8 min, 4–6 month payback demonstrates regulated industry viability.
— Azure Build 2026 announcements for Foundry IQ: agentic retrieval quality improvements (evidence recall +46–54%, token cost −34%, answer quality +8–20%), new knowledge sources (Work IQ, Fabric Ontology) enabling cross-source retrieval.
— Vendor GA documentation providing systematic methodology for optimizing domain-specific RAG retrieval quality; emphasizes metadata filtering (>90% search space reduction), hybrid search, and reranking as domain-specific optimization priorities.
— Production domain-specific RAG: 50K Russian corporate documents improved Top-1 from 62% to 88% via hybrid search (BM25+vector RRF) and cross-encoder reranking; demonstrates retrieval architecture ladder and domain-specific optimization strategies.
— Stanford research shows RAG reduces legal AI hallucinations from 58-82% to 17-34%, demonstrating RAG's significance but exposing persistent limitations in regulated domains; critical negative signal for high-stakes applications.
— Goodyear production agentic RAG deployment over 2,100 academic papers (24K chunks) with evidence sufficiency scoring and cross-corpus academic search; demonstrates integration of internal knowledge graphs with external academic databases.
— Official Azure AI Search agentic retrieval documentation (public preview 2026-06-01) designed for cross-corpus question answering via LLM-driven query decomposition and parallel execution; production infrastructure for complex domain-specific QA.
— Financial RAG production analysis quantifying failure modes: corpus audit yield +11% vs algorithmic improvements +5%; recall improved 0.83→0.94; demonstrates data quality bottleneck and bottleneck shift to generation once retrieval improves.
— Five named Fortune 500 RAG failure post-mortems ($12.2M trading loss, $4.1M legal settlement, $2.3M manufacturing loss) document temporal grounding and retrieval necessity; demonstrates RAG criticality in specialized domains and cost of removal.
— Novel framework for domain-specific RAG evaluation with six diagnostic metrics and 200 curated QA pairs; evaluates context-scaling behavior across models revealing attention dilution as primary degradation mechanism in specialized corpora.
— Peer-reviewed medical benchmark (JMIR) with 511 MSI cancer questions; RAG as most effective intervention; demonstrates critical success factors: retrieval precision and knowledge base quality determine system safety.
— Peer-reviewed materials science RAG achieving 85.6% accuracy vs 21.3% baseline (4x improvement) via structured knowledge + query reformulation on water-splitting catalysis domain; production-scale deployment with quantified ROI.
— Clinical RAG benchmark (OGCaReBench) on 513 expert-validated questions shows 56% baseline to 82% with retrieval augmentation; demonstrates RAG necessity for medical domain evidence-grounded reasoning beyond guidelines.
— Bilingual (French/English) legal RAG benchmark (ClaimRAG-LAW) targeting experts and non-experts with fine-grained evaluation framework and hallucination analysis; addresses critical gap in domain-specific legal AI evaluation.
— Cross-domain medical knowledge alignment via Query-Conditioned Entity Alignment (QCEA) improves evidence retrieval and answer accuracy in heterogeneous domains (Traditional Chinese vs Western Medicine); demonstrates semantic integration requirement for cross-corpus medical RAG.
— Empirical medical domain case study across 6 specialties: hybrid BM25+dense retrieval achieves Recall@5 0.86 vs BM25-only 0.71; complementary strengths confirmed with statistical significance (McNemar p<0.001).
— SIGIR 2026: passage structure and quality directly impact domain-specific QA for inferential reasoning; convergence metric for constructing reasoning-optimal passages outperforms cosine similarity selection across six LLM architectures.
— SIGIR 2026 benchmark for category-aware ranking across heterogeneous sources (Publications, Data, Variables, Tools); addresses production requirement of unified retrieval interfaces over diverse domain-specific data types.
— AAAI 2026 empirical study on automotive domain-specific QA: RAG proves most cost-efficient adaptation for both premium and open-source models, with detailed cost-quality tradeoff analysis for industrial closed-domain datasets.
— Independent audit of 50 live deployments documenting 7 production failure modes (context gaslighting 76% finance, citation fabrication 81% legal) across adversarial testing; critical negative signal confirming production readiness demands meticulous deployment discipline.
— Peer-reviewed clinical benchmark: domain context variables (corpus type, query format) explain 49% variance in retrieval performance vs 47.6% for model choice; model rankings unstable across clinical corpora, confirming domain-specific validation mandatory.
— Novel agentic retrieval evaluation framework (BRIGHT-Pro, RTriever-Synth, RTriever-4B) for multi-aspect evidence portfolios across search steps, addressing leading-edge requirement to support reasoning-intensive cross-corpus QA beyond single-shot retrieval.
— Peer-reviewed empirical comparison of 5 retrieval strategies (Dense, Hybrid, Cross-Encoder, Multi-Query, MMR) on BioASQ benchmark showing cross-encoder reranking achieves 0.827 composite score (0.852 contextual precision), establishing domain-specific retrieval strategy selection guidance.
— Real domain-specific RAG deployment (LLC formation, logistics, AI automation with 4,290 vectors) suffered silent embedding model mismatch (ONNX-quantized indexing vs Hugging Face API queries) below similarity threshold. Demonstrates embedding consistency as critical infrastructure requirement.
— Benchmark revealing structural/semantic errors from high-accuracy OCR (11 document types tested) still cause RAG retrieval failures. Challenges assumption that OCR accuracy alone ensures downstream RAG viability in document-heavy domains.
— Full production deployment of domain-specific semantic search at children's hospital serving 1.68M patients across 166M clinical notes. Qwen3 embeddings achieved 94.6% clinical QA accuracy with 237ms latency and $4K/month cost, confirming health-system-scale viability.
— Production reranker models (0.8B-9B) generating relevance + contribution statement + evidence passage for agentic/cross-corpus RAG, improving BEIR scores. Addresses cross-corpus context precision requirement.
— Legal contracts RAG production deployment (900+ documents): naive 512-token chunking achieved 0.61 precision; parent-child chunking + hybrid search + reranking reached 0.87, demonstrating domain-specific document structure as primary architectural driver.
— RAG framework for high-redundancy enterprise domains (finance, legal, patents). Quantifies robustness degradation on enterprise corpora: dense retrievers drop from 66.4% to 5.0-27.9% recall across domain-specific hops, establishing evaluation methodology for specialized domains.
— Production analysis of hubness phenomenon in vector retrieval where high-dimensional geometry causes disproportionate chunk retrieval. Provides diagnostic framework (Gini coefficient, k-occurrence metrics) and stacked mitigations (MMR, reranking) for domain-specific RAG systems.
— Domain-specific RAG benchmark for sensitive domains with quantified deployment results: domain adaptation fine-tuning achieved 26% task success improvement and 47% hallucination reduction on defense document QA.
— Peer-reviewed production RAG deployment for legal domain with real metrics: 184,895 audited answers, 81.7% legislation reference resolution, 47.1% jurisprudence resolution, 6.5% hallucination correction rate.
— Uber's production case study: Genie on-call copilot achieved 27% relative improvement in acceptable answers and 60% reduction in incorrect advice through enhanced agentic RAG for domain-specific (security/privacy) policies.
— Practitioner analysis identifying document quality as critical RAG failure mode. Reports specific metric: document quality scoring improved search accuracy from 62% to 89% with no embedding/retrieval changes.
— SIGIR 2026 paper introducing ConflictQA benchmark for cross-source knowledge conflicts (text vs. knowledge graphs) and XoT reasoning framework—directly addresses cross-corpus QA challenge.
— Domain-specific RAG case study in medical/biomedical domain using specialized embedding models (MedCPT) and domain-specific evaluation. Tests retrieval augmentation on biomedical corpora with measured grounding and answerability metrics, demonstrating domain adaptation of RAG architectures for knowledge-intensive medical QA.
— Production-focused guidance on embedding selection for domain-specific RAG. Highlights MTEB benchmark limitations and index drift risks from silent model versioning changes.
— Peer-reviewed systematic evaluation of domain-specific RAG pipeline components for medical QA, showing empirical optimization of retrieval strategies within domain constraints.
— Empirical benchmark data showing domain-specific fine-tuning improvements: medical +29% (48%→62%), code +34% (44%→59%), legal +24% (46%→57%). Direct signal that general models fail on specialized domains.
— Microsoft domain-adapted embedding model (0.6B params) trained for legal, finance, healthcare, code; outperforms voyage-3.5 (0.7241 vs 0.7139) despite tiny parameter count, signals ecosystem shift toward specialized embeddings.
— Production RAG maintenance challenge: vector drift causes silent retrieval degradation from embedding model version mismatch, incremental content updates without re-embedding, inconsistent chunking strategies.
— Production RAG failures across 12+ systems (fintech, healthcare, SaaS): naive chunking (54%→81% precision fix), wrong embeddings (domain-adapted fixes), missing re-ranker, context compression; demonstrates specialization necessity.
— Three production domain-specific RAG systems: multilingual furniture (10K+ products, 85%+ cross-language precision), B2B configurators with domain logic, KG integration; pragmatic analysis of domain-specific tradeoffs vs context stuffing.
— Domain-specific RAG benchmark on financial documents (23,088 queries): hybrid retrieval + neural reranking achieves 0.816 Recall@5; BM25 outperforms dense retrieval in financial domain, challenging semantic-search universality.
— EACL 2026 StackRepoQA benchmark: domain-specific code cross-corpus QA on 134 Java projects (1,318 questions); structural signals improve RAG but overall accuracy limited, demonstrating repository-scale comprehension complexity.
— Azure AI Search 2025-2026 evolution: November 2025 GA for agentic retrieval, answer synthesis, semantic ranker on free tier; signals mainstream vendor recognition of domain-specific RAG as core capability.
— Peer-reviewed study on domain-specific RAG for AI policy using AGORA corpus (947 documents): domain fine-tuning improves retrieval metrics but not end-to-end QA; stronger retrieval paradoxically increases confident hallucinations.
— Healthcare domain-specific RAG: fine-tuned multilingual embeddings on 400K clinical documents (163K patients) achieved mAP@100 0.27 vs. baselines 0.14-0.11, demonstrating clinical embedding specialization necessity.
— Competition-validated domain-specific legal RAG system achieves 0.935 MIR-E score on QIAS 2026 leaderboard via rule-grounded synthesis, hybrid retrieval, neural reranking, schema-constrained generation on Islamic law.
— Recent EACL conference paper identifying fundamental misalignment in RAG evaluation: traditional IR metrics assume sequential human examination, but LLMs process all documents holistically; proposes UDCG metric improving correlation by 36%.
— Direct match: peer-reviewed research on RAG for knowledge-intensive multilingual QA. Proposes solutions (CrossRAG) for cross-corpus consistency challenges, published in top-tier venue (EACL 2026 Findings).
— Peer-reviewed research from Meta AI and Google DeepMind proposing ART (unsupervised dense retriever training) for QA retrieval. Advances unsupervised methods for domain-specific question answering systems.
— Production implementation of domain-specific RAG across regulated domains (medical, legal, financial) with substantial empirical improvements and enterprise governance features demonstrating leading-edge maturity.
— Comprehensive market research synthesizing 85+ sources on RAG adoption rates, ROI metrics, failure modes, and architectural evolution; signals leading-edge maturity with named case studies and cost-benefit data.
— Domain-specific RAG adaptation framework for specialized fields (science, medicine). Proposes self-training approach equipping LLMs with joint question answering and question generation capabilities for domain-specific knowledge systems.
— Real deployment of multi-stage retrieval system for cross-corpus agent memory achieving 96-100% single-session accuracy; demonstrates domain-specific retrieval at scale (115K+ tokens) with measured model comparisons.
— Production case study with quantified outcomes. Domain-specific RAG system deployed at scale, reducing hallucination by 72% and serving millions of interactions. GraphRAG architecture for relational knowledge.
— Critical analysis citing real-world failures: healthcare harmful guidance from outdated knowledge base, finance compliance violation, telecom $2.3M service credits for incorrect information—demonstrates deployment risks in high-stakes domains.
— Azure AI Search February 2026 release adds agentic retrieval portal support with new knowledge sources (OneLake, SharePoint, Web) and retrieval options, signaling vendor infrastructure evolution for domain-specific RAG.
— Practitioner Azure MVP deployed multiple domain-specific RAG systems (product catalog, resume search) with embeddings and Azure AI Search, demonstrating 11-month real-world implementation journey.
— Official Azure documentation detailing agentic and classic RAG patterns with guidance on solving domain-specific challenges including query understanding, multi-source access, and token constraints.
— Research addressing RAG security vulnerabilities via sparse document attention mechanism; shows defenses substantially outperform standard approaches, indicating maturity in security-critical domain deployments.
— Research reveals RAG and reranking fail at inference-based reasoning on QUIT dataset; current QA pipelines not ready for indirect evidence reasoning, exposing fundamental capability limitations.
— Research benchmark showing standard RAG systems collapse on 10M token corpora (1.20% accuracy vs. 55.85% at 128K), while memory-augmented agents demonstrate resilience; indicates architectural limits for large-scale cross-corpus retrieval.
— Domain-specific RAG evaluated on 663 biologics package inserts and 206 monoclonal antibodies; retriever showed inaccuracies despite coherent LLM outputs, highlighting retriever-generation gaps in specialized medical domains.
— Advanced RAG with cross-encoder re-ranking on CDC policy documents achieves 0.797 faithfulness vs. 0.347 baseline; demonstrates two-stage retrieval effectiveness and document segmentation constraints for multi-step domain-specific reasoning.
— Critical assessment of RAG limitations in high-stakes domains: retrieval quality, evolved hallucinations, context constraints, freshness issues, and evaluation complexity; cites healthcare risks and 'lost in the middle' phenomenon.
— Framework for cross-domain RAG with domain routing and query rewriting shows improved accuracy and citation consistency across academic, tourism, and finance domains; addresses cross-domain generalization via transfer-oriented training.
— Critical analysis of RAG architecture bifurcation: caching for small corpora (<1M tokens) with 15-20% accuracy gains vs. agentic hypergraphs for complex domains; signals emerging consensus that standard RAG is obsolete for modern workloads.
— Framework for rigorous domain-specific RAG evaluation using synthetic QA generation and LLM-as-a-judge metrics; demonstrates evaluation challenges across military, cybersecurity, and engineering domains with context-dependent performance.
— Real deployment failure: GPT-4.1 agent with Azure AI Search fails to consistently invoke search tool, exposing challenges in enforcing retrieval behavior and potential tool schema/configuration issues.
— Industry analysis reporting that RAG powers 60% of production AI applications in 2025; signals broad enterprise adoption of RAG across customer support, knowledge bases, and internal systems.
— Production RAG deployment for SalesWorx product documentation using LlamaIndex, Qdrant vector DB, and multi-agent orchestration with 262K context; demonstrates real-world implementation architecture for domain-specific knowledge retrieval.
— Configuration pitfall in Azure AI Search: Text Splitting skill silently failed due to incorrect context property; highlights specific implementation barriers and need for precise configuration in production domain-specific RAG systems.
— Healthcare RAG study: scope-matched knowledge graphs deliver consistent gains, while indiscriminate graph unions introduce distractors; establishes principle that precision-first, scope-focused KG-RAG outperforms breadth-first approaches.
— RAGen framework for generating domain-specific QA training data without labeled datasets; enables RAG adaptation via semantic chunking and hierarchical concept extraction, addressing key barrier of limited domain-specific training data.
— RANLP 2025: domain-specific QA framework integrating KGs with RAG achieves 13% accuracy improvement on financial datasets; demonstrates automated KG construction and question decomposition for domain-specific retrieval.
— Study across diverse corpora (Wikipedia, PubMed, GitHub, StackExchange) shows RAG helps small LLMs (+22.87%) but provides minimal gains for large models; reveals that corpus routing remains unsolved with LLMs poor at dynamic source selection.
— Critical analysis documenting domain-specific RAG accuracy plateau at 75% (25% error rate) in production; advocates graph-based semantic layers over vector-only RAG, citing >95% success on policy document domain vs. vector-only underperformance.
— Coxwave Align production deployment: domain-specific embedding fine-tuning via NeMo Curator achieved 12% RAG accuracy improvement and 6x training speedup, demonstrating practical performance gains in real-world analytics platform.
— SIGIR 2025 LiveRAG competition winner: knowledge-aware reranking RAG pipeline achieving first place on 15M-document FineWeb corpus evaluation, demonstrating cross-source QA capability and retrieval effectiveness at scale.
— Practitioner analysis of multi-document long-context QA evaluation challenges: information overload, positional variance, hallucinations at scale; identifies tension between faithfulness and helpfulness in domain-specific deployments.
— Case narrative illustrating domain-specific RAG deployment friction: SaaS knowledge base search returning 15 irrelevant articles for specific configuration queries; highlights practical effectiveness gaps in production knowledge-base RAG systems.
— Enterprise knowledge QA benchmark spanning diverse corporate document types (product releases, technical blogs, financial reports); validates domain-specific RAG effectiveness for real enterprise knowledge bases with frequently updated content.
— NAACL 2025 research on domain adaptation: SimRAG self-training approach enabling LLMs to perform QA and generate questions for science/medicine domains, addressing distribution shift and limited access to specialized data.
— Evaluates KG-RAG failure modes when knowledge graphs have missing information; demonstrates practical limitations of domain-specific KG-based retrieval in real-world scenarios where corpus completeness cannot be guaranteed.
— AWS GA announcement (March 2025) for RAG Evaluation enabling domain-specific application assessment via context relevance, coverage, faithfulness, and responsible AI metrics; signals major cloud vendor platform maturity.
— FacetTrak field service management platform deployed Azure AI Search with semantic search for domain-specific ticket retrieval; production rollout with automated migration and daily indexing, improving search experience in operational domain.
— Framework for domain-specific RAG optimization on finance (SEC 10-K), biomedical (PubMed), and cybersecurity (APT) corpora; smaller chunks (<10 tokens) improve precision by 31-42%, with domain-specific embedding yielding 22% variance in optimal sizing.
— Large-scale empirical evaluation of RAG on 20,000 FAQ queries with 400,000 KB entries; RAG showed knowledge grounding but DoRA outperformed on accuracy (90.1%) and latency (110ms), revealing limitations in accuracy-critical domains like healthcare and finance.
— Practitioner analysis of real-world RAG implementation barriers: poor retrieval in healthcare, semantic mismatches in finance, outdated knowledge bases; highlights critical adoption obstacles in high-stakes domains requiring extensive fine-tuning and domain-specific indexing.
— CIKM 2025 research on domain-specific RAG for automotive industry (crash test docs); pilot deployment showed +1.79 factual correctness, +1.33 informativeness vs. baseline, confirming real-world adoption in manufacturing domain.
— Critical analysis documents nine RAG limitations: retrieval quality dependency, latency/scalability, context limits, transparency gaps, bias vulnerability, maintenance costs, ethical concerns, and domain-specific brittleness.
— Microsoft releases agentic retrieval features, multimodal RAG support, and knowledge base source expansions; demonstrates continued platform investment in domain-specific RAG infrastructure.
— Experience report identifies seven critical failure points across three case studies (cognitive reviewer, AI tutor, biomedical QA); highlights that RAG validation is only feasible in operation and robustness evolves rather than being designed in.
— Case study builds RAG system over 1,800+ web pages with hybrid BM25/FAISS retrieval; F1 improves from 5.45% to 42.21%, demonstrating practical domain-specific application.
— Domain adaptation on hotel customer service domain shows fine-tuning improves QA performance and significantly reduces hallucinations across all evaluated RAG architectures.
— Case study of domain-specific RAG for ophthalmology with 70,000 clinical documents; RAG reduced hallucinations to 26.7% (vs. 45.3% without), increased evidence accuracy to 54.5%, demonstrating real-world medical domain deployment.
— Analysis of seven production RAG failure points (missing content, incorrect specificity, ranking, context, format, extraction, output); signals deployment barriers requiring extensive testing and fine-tuning beyond basic setup.
— IEEE BESC 2024 conference paper on domain-specific RAG applied to niche domain (vocal training); demonstrates knowledge segmentation and semantic similarity approaches for highly specialized corpora.
— SMART-SLIC framework integrates RAG with domain-specific knowledge graphs and vector stores; builds specialized corpora without LLM hallucinations, enabling source attribution and reducing fine-tuning burden.
— Identifies eight critical failure points in KG-based RAG; proposes Mindful-RAG framework targeting intent-based retrieval and contextual alignment to address fundamental retrieval brittleness in domain-specific systems.
— EuroPython 2024 talk examining RAG limitations and failure modes across domains; argues domain-specific RAG is not one-size-fits-all, with vastly different quality and transparency requirements by use case.
— Business intelligence deployment of conversational document search using Azure AI Search with hybrid retrieval and OpenAI embeddings; real production system for domain-specific RAG.
— Practitioner report of Azure AI Search vector search limitation: returns empty results beyond 1000 items, revealing scalability constraint affecting large-corpus domain-specific RAG deployments.
— College enrollment domain benchmark identifies six required RAG capabilities; finds existing LLMs struggle with domain-specific questions, confirming need for specialised retrieval-augmented approaches.
— Adobe products QA system shows fine-tuning retrievers reduces hallucinations and improves generation; demonstrates production-stage domain-specific RAG deployment in corporate environment.
— Financial filings RAG study shows fine-tuning embedding models and LLMs improves accuracy; iterative reasoning further boosts performance toward human-expert quality on domain-specific corpora.
— LFRQA benchmark across 7 domains finds only 41.3% of top LLM RAG answers preferred over human references, revealing cross-domain robustness as persistent challenge limiting generalization.
— Azure AI Search announces 11x vector index capacity increase, 6x storage increase, 2x throughput improvements at no cost, with Fortune 500 adopters (OpenAI, KPMG, PETRONAS) confirming production deployment at scale.
— Domain-specific RAG techniques for financial documents: chunking, query expansion, metadata annotation, re-ranking, and embedding fine-tuning for 10-Ks and earnings transcripts.
— End-to-end system design for domain-specific RAG on private knowledge bases (CMU case); demonstrates approach to hallucination mitigation but reveals limitations of fine-tuning with small, skewed datasets.
— Comprehensive 353-paper survey providing unified taxonomy of RAG foundations, enhancements, and applications; signals research consolidation and maturity in Q1 2024.
— Enterprise RAG deployment (SharePoint, 1100 files) shows inconsistent answer retrieval and parameter-tuning difficulties; signals adoption barrier where increasing retrieval volume reduces accuracy.
— Community-curated RAG research collection (1.8k stars) supporting the AIGC survey; demonstrates sustained open-source engagement and practitioner interest in RAG taxonomies.
— Deployment barrier: indexing large PDF documents fails with vectorization timeouts; reveals token limits (8,191 tokens/6k words) and need for document chunking in production systems.
— Real-world deployment challenge: Azure AI Search RAG system returning inconsistent results across similar queries; highlights fundamental retrieval accuracy barriers in early 2024.
— Comprehensive survey consolidating RAG paradigms (Naive, Advanced, Modular), retrieval/generation/augmentation components, and evaluation frameworks; signals academic maturation and recognition of domain-specific information integration benefits.
— Framework for generating domain-specific, multimodal, multi-hop QA datasets to evaluate RAG systems; addresses gap in evaluation of specialized technical documents with >2.3 average reasoning hops.
— EMNLP 2023 study showing corpus incompleteness limits retrieval-based QA; LLM-generated passages outperform retrieved ones by 10-13%, highlighting fundamental limitations of corpus-only RAG without generation augmentation.
— Hacker News practitioner discussion highlighting RAG challenges: embedding failures on domain terminology, testing complexity, tool immaturity (LangChain criticism), expensive vector databases; signals implementation barriers.
— ACL 2023 benchmark across 8 domains (finance, medicine, law) showing ODQA models trained on Wikipedia fail to generalize; demonstrates significant gap in cross-domain robustness for domain-specific RAG.
— IBM Research's open-source toolkit supporting domain-specific QA with retrieval and reading comprehension; enables replication of state-of-the-art methods and front-end applications.
— Discussion of the foundational RAG framework combining parametric and non-parametric memory for knowledge-intensive tasks, enabling QA over external corpora.
— ACL 2023 paper demonstrating task-aware specialization in dense retrieval models for efficient open-domain QA, advancing domain-specific retrieval.
— Empirical study examining KG-RAG effectiveness across diverse domains and scenarios, evaluating knowledge graph quality impact on RAG performance.
— Technical analysis of RAG hybrid architecture combining pre-trained parametric models with explicit non-parametric memory for improved knowledge retrieval.
— DeepPavlov's KBQA production system enables domain-specific QA over structured knowledge bases like Wikidata, supporting complex multi-step queries.
— Research on neural retrievers using cross-architecture knowledge distillation to improve open-domain QA performance, advancing dense retrieval methods.