\n
For researchers, analysts, and security professionals alike, the ability to\u00a0quickly and accurately retrieve relevant information\u00a0is critical. Yet, as our information landscape grows, so do the challenges of traditional search methods.<\/p>\n
The Cisco Foundation AI team\u00a0introduces a novel approach to information retrieval designed to tackle the shortcomings of current search.<\/p>\n
The Challenge with Current Search<\/h2>\n Often, when we search for information,\u00a0especially for\u00a0complex topics, our\u00a0initial\u00a0queries might not hit the mark. Traditional search engines, while powerful, typically\u00a0operate\u00a0on a \u201cone-shot\u201d principle: you ask a question, and it gives you results. If those results\u00a0aren\u2019t\u00a0quite right,\u00a0it\u2019s\u00a0up to you to reformulate your query and try again. This process can be inefficient and frustrating, particularly when dealing with nuanced or multi-faceted information needs.<\/p>\n
LLMs offer semantic understanding, but they can be computationally expensive and not always ideal for the iterative, exploratory nature of complex searches.\u00a0Existing methods for query rewriting or decomposition often commit to a search plan too early, causing the retrieval process to become trapped in an incorrect search space and miss relevant information.<\/p>\n
Foundation AI\u2019s\u00a0Adaptive Approach<\/h2>\n The Foundation AI\u00a0approach to search\u00a0addresses these limitations by making the retrieval process itself adaptive and intelligent.\u00a0Instead of a static, one-and-done query,\u202fthe framework\u00a0enables models to learn how to\u202fsearch<\/em>\u202fiteratively, much like a human investigator would.\u00a0This is done using a series of techniques: synthetic trajectory generation to create diverse search behaviors, supervised fine-tuning to\u00a0establish\u00a0the scaffolding for multi-turn search, reinforcement learning (GRPO) to refine search behavior, and finally inference time beam search to exploit the learned self-reflection capabilities.<\/p>\nAt its core,\u202four framework\u00a0empowers compact models (from 350M \u2013 1.2B parameters) to:<\/p>\n
\nLearn diverse search strategies:<\/strong>\u202fThrough a process of\u00a0observing\u00a0and learning from various search behaviors,\u00a0the framework\u00a0models understand how to\u00a0approach\u00a0different types\u00a0of queries.<\/li>\nRefine queries based on feedback:<\/strong>\u202fThe system learns to adjust its search queries dynamically, incorporating insights from previously retrieved documents.<\/li>\nStrategically backtrack:<\/strong>\u202fA critical capability is knowing when to abandon an unfruitful path and explore alternative search directions, preventing the \u201crevolving loops\u201d seen in less adaptive systems.<\/li>\n<\/ul>\nTogether, these abilities allow\u202four search framework\u202fto conduct a multi-turn\u00a0\u201cconversation\u201d with the information it retrieves,\u00a0reflect on\u00a0intermediate results,\u00a0and\u00a0adapt\u00a0its strategy to zero in on the most relevant evidence. The figure below compares some of the existing approaches discussed with that of\u202fthe Foundation AI\u00a0team\u2019s\u00a0approaches.<\/p>\n
\n
Figure 1: Overview of\u00a0framework<\/figcaption><\/figure>\n<\/div>\nWe illustrate two established query reformulation baselines alongside our\u00a0proposed\u202fframework on an example from the FEVER dataset. While query\u00a0decomposition fails without corpus feedback and query rewriting yields static\u00a0reformulations that ignore retrieval results,\u202fthe Foundation AI\u00a0framework\u00a0performs tree-based exploration with\u00a0structured reasoning spans, revising its strategy as it incorporates contradictory evidence\u00a0and shifts from valley- to mountain-focused queries-effectively backtracking, refining, and\u00a0exploring to recover relevant evidence.<\/p>\n
Results<\/h2>\n We evaluated\u202four\u00a0approach\u00a0across two challenging benchmark suites that test both retrieval precision and reasoning depth: the BEIR benchmark for classic and multi-hop information retrieval, and the BRIGHT benchmark for reasoning-intensive search spanning scientific, technical, and analytical domains.<\/p>\n
Despite being up to 400\u00d7 smaller than the large language models it was compared against,\u202four smaller custom models\u00a0used in the tests\u202fconsistently performed at or above par:<\/p>\n
\nOn BEIR datasets such as\u00a0SciFact, FEVER,\u00a0HotpotQA, and\u00a0NFCorpus,\u202fthe Foundation AI\u00a0large (1.2B)\u00a0model\u00a0achieved\u202f77.6%<\/strong>\u202f\u00a0nDCG@10 on\u00a0SciFact\u00a0and\u00a0\u202f63.2%<\/strong>\u202fnDCG@10\u00a0on\u00a0NFCorpus, surpassing prior retrievers and approaching GPT-4-class performance, while\u00a0maintaining\u00a0strong scores on FEVER (65.3%)\u00a0and\u00a0HotpotQA\u00a0(71.6%).<\/li>\nOn BRIGHT, we\u00a0achieved a macro-average nDCG@10 of\u202f25.2%<\/strong>, outperforming large proprietary models like GPT-4.1 (22.1%) across 12 diverse domains, from economics and psychology to robotics and mathematics.<\/li>\n<\/ul>\nThese results\u00a0demonstrate\u00a0that\u202flearned adaptive search strategies,<\/strong>\u202fnot just model scale, drive retrieval performance.<\/p>\nReal-world\u00a0Application:\u00a0Security Search<\/h2>\n The implications of such an adaptive retrieval system\u00a0reach\u00a0across domains, especially in security:<\/p>\n
\nEnhanced Threat Intelligence Analysis<\/strong>:\u202fSecurity analysts are constantly sifting through massive volumes of threat reports, vulnerability databases, and incident data. The\u00a0framework\u2019s\u00a0ability to handle complex, evolving queries and backtrack from dead ends means it can more effectively uncover subtle connections between disparate pieces of intelligence,\u00a0identifying\u00a0emerging threats or attack patterns that a static search might miss.<\/li>\nFaster Incident Response<\/strong>:\u202fWhen a security incident takes place, responders need to quickly\u00a0locate\u00a0relevant logs, network traffic data, and security policies.\u202fAccelerate\u00a0this by adaptively searching through diverse data sources, refining\u00a0queries as new evidence\u00a0emerges\u00a0from the incident, and helping to pinpoint the root cause or affected systems\u00a0faster.<\/li>\nProactive Vulnerability Research<\/strong>:\u202fSecurity researchers can use\u00a0the framework to\u00a0explore code repositories, technical forums, and security advisories to\u00a0identify\u00a0potential vulnerabilities in systems. Its adaptive nature allows it to follow complex chains of dependencies or exploit techniques, leading to more comprehensive vulnerability discovery.<\/li>\n<\/ul>\nThe Future of Search is Adaptive<\/h2>\n Our research\u00a0shows that retrieval intelligence is not a function of scale but of strategy. By combining synthetic data, reinforcement learning, and intelligent search algorithms, compact models can achieve powerful adaptive capabilities. This means more efficient,\u00a0cost-effective, and robust information retrieval systems that can\u00a0truly understand\u00a0and adapt to the complexities of human information needs.\u00a0<\/p>\n
If\u00a0you\u2019re\u00a0interested in learning more, you can read the full research paper\u00a0\u202fhere on arXiv.<\/p>\n
Learn more about the research we do and sign up for updates at the\u00a0Cisco Foundation AI team\u00a0website.<\/p>\n
We\u2019d love to hear what you think! Ask a question and stay connected with Cisco Security on social media.<\/em><\/p>\nCisco Security Social Media<\/mark><\/strong><\/p>\nLinkedInX<\/a><\/p>\n<\/p><\/div>\n
![\"Retail](\"https:\/\/dmsretail.com\/RetailNews\/wp-content\/uploads\/2022\/05\/RETAIL-ONLINE-TRAINING-728-X-90.png\")
<\/a><\/p>![\"Search](\"https:\/\/blogs.cisco.com\/gcs\/ciscoblogs\/1\/2026\/01\/search-framework-graphic-1024x542.webp\")