{"id":97182,"date":"2026-09-14T13:38:48","date_gmt":"2026-09-14T10:38:48","guid":{"rendered":"https:\/\/twodots.gr\/?p=97182"},"modified":"2026-09-14T13:38:49","modified_gmt":"2026-09-14T10:38:49","slug":"schemarouter-ai-agents-field-aware-rag","status":"publish","type":"post","link":"https:\/\/twodots.gr\/en\/schemarouter-ai-agents-field-aware-rag\/","title":{"rendered":"SchemaRouter: How AI Agents Find the Right Data Without Unnecessary Context"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>Answer first:<\/strong> SchemaRouter shows that an AI agent does not need to view the entire list of APIs or retrieve all available fields to respond correctly. All that is needed is a small, executable plan that selects the right tools, retains the critical fields, and passes along the units, origin, and license information with them.<\/p>\n<p>In the study\u2019s benchmark, the approach achieved an answer accuracy of 0.71 with 227 retrieved-context tokens, compared to 2,066 tokens for the fetch-all method, which had an accuracy of 0.70. The result pertains to 110 materials science queries and does not demonstrate universal superiority. However, it offers a specific architectural lesson for agents connected to CRM, analytics, e-commerce catalogs, databases, and external APIs.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Contents<\/div>\n<ul>\n<li><a href=\"#kryfo-kostos-agentic-rag\">The Hidden Costs Behind an Agentic RAG Workflow<\/a><\/li>\n<li><a href=\"#prompt-all-over-fetch-under-fetch\">Why \"prompt-all,\" \"over-fetch,\" and \"under-fetch\" cause different problems<\/a><\/li>\n<li><a href=\"#schema-graph-leitourgikos-chartis\">The schema graph converts APIs into a functional map<\/a><\/li>\n<li><a href=\"#ektelesimo-tool-plan\">How an executable tool plan is generated<\/a><\/li>\n<li><a href=\"#benchmark-110-erotimata\">What Does the 110-Question Benchmark Really Measure?<\/a><\/li>\n<li><a href=\"#apotelesmata-ligotero-context\">The results: competitive accuracy with much less context<\/a><\/li>\n<li><a href=\"#elachista-pedia-lathos-kpi\">Why the minimum number of fields is the wrong KPI<\/a><\/li>\n<li><a href=\"#provenance-adeies-asfaleia\">Provenance, Licenses, and Data Security<\/a><\/li>\n<li><a href=\"#oria-schemarouter\">Where Does SchemaRouter Still Fall Short?<\/a><\/li>\n<li><a href=\"#crm-ecommerce-marketing-agents\">What does this mean for CRM, e-commerce, and marketing agents?<\/a><\/li>\n<li><a href=\"#pilot-epta-vimata\">A seven-step pilot for field-aware routing<\/a><\/li>\n<li><a href=\"#mikrotero-eparkes-synolo\">The Essence: The Smallest Sufficient and Verifiable Set<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"kryfo-kostos-agentic-rag\">The Hidden Costs Behind an Agentic RAG Workflow<\/h2>\n<p>The more tools an AI agent acquires, the more difficult a seemingly simple decision becomes: which tool should it call, with what parameters, and exactly which fields does it need to return? In a simple chatbot, the model responds based on the context it is given. In an agentic RAG system, it must first decide where to search and then convert that choice into a valid call.<\/p>\n<p>A query may require a combination of an internal database, a vector store, a graph store, and an external API. The choice isn\u2019t limited to the name of the tool: each endpoint has different required parameters, response fields, units, sources, and usage restrictions. If any of these are missing, the plan may be semantically relevant but technically unfeasible.<\/p>\n<p>The Research Paper <em>SchemaRouter: Field-Aware Tool Routing for Efficient Heterogeneous Agentic RAG<\/em>, published on arXiv on July 3, 2026, proposes an intermediate routing layer. A small LLM extracts the intent and key concepts; then, deterministic rules based on a schema graph select tools and fields. This distinction is critical because the <a href=\"https:\/\/twodots.gr\/ai-agents-poios-ftaiei-otan-aftomatopoiisi-apotygchanei\/\">Diagnosing an AI agent failure<\/a> It requires us to know whether the problem was caused by the interpretation, the plan, the tool call, or the downstream reader.<\/p>\n<h2 id=\"prompt-all-over-fetch-under-fetch\">Why \"prompt-all,\" \"over-fetch,\" and \"under-fetch\" cause different problems<\/h2>\n<p>The Strategy <strong>prompt-all<\/strong> It places all tool and field descriptions within the routing prompt. This may be effective with a robust model, but the schema is repeated with every call and grows along with the inventory. In the study, the catalog contained 19 tools and 166 response fields. The schema segment alone had 5,042 characters, or approximately 1,260 tokens, based on the simplified assumption of four characters per token.<\/p>\n<p>The authors estimate an increase of approximately 66 routing-prompt tokens per additional tool: about 6,600 tokens for 100 tools and 33,000 for 500. This is a projection based on the inventory of 19 tools, not a measurement from deployments of this scale. Prompt caching can reduce the input cost of a fixed schema, but it does not eliminate the need to pay attention to the full prefix, nor does it address changes to tools, versions, or permissions.<\/p>\n<p>The <strong>over-fetching<\/strong> This is a different problem: the agent calls more tools or returns more fields than are needed. This increases the payload, the context, and the likelihood that sensitive or conflicting data will reach the reader. The <strong>under-fetching<\/strong> leaves out a crucial area and may lead to a convincing but incomplete answer. Protection against <a href=\"https:\/\/twodots.gr\/ai-agents-indirect-prompt-injection-epixeiriseis\/\">Indirect prompt injection in external sources<\/a> It remains a separate check; a lower recovery rate limits the coverage, but does not prove that the content is reliable.<\/p>\n<div class=\"td-comparison td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid--three\">\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">Prompt-all<\/p>\n<p>It provides the model with the entire tool schema on every routing call. It offers broad visibility, but the prompt grows as the inventory expands.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Complete schema<\/span><span class=\"td-badge\">High routing cost<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">Vector-only<\/p>\n<p>It finds semantically similar tools, but may fall short in terms of required parameters, response fields, and executability.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Similarity<\/span><span class=\"td-badge\">Risk of under-fetch<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">SchemaRouter<\/p>\n<p>It uses a small LLM for understanding and deterministic field projection to generate valid, metadata-aware tool plans.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Field-aware<\/span><span class=\"td-badge\">Executable plan<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"schema-graph-leitourgikos-chartis\">The schema graph converts APIs into a functional map<\/h2>\n<p>SchemaRouter represents Tool, Endpoint, Parameter, ResponseField, DomainConcept, Unit, SourceType, LicensePolicy, FieldGroup, and Intent as nodes. Edges connect endpoints to parameters and fields, fields to units, sources, and concepts, and natural language expressions to canonical field names. The system does not merely store textual descriptions; it stores the relationships that determine which requests are valid and what each response means.<\/p>\n<p>The concept is similar to schema linking in text-to-SQL, but it is adapted to an environment with many heterogeneous tools. It does not merely map words to database columns. It links queries to fields across multiple APIs and maintains metadata that a downstream reader cannot safely reconstruct from parametric knowledge.<\/p>\n<p>The production graph that inspired the study is proprietary. The paper references a frozen JSON export, an evaluation harness, and a repository for reproducing the results. However, when checked on September 14, 2026, the cited GitHub link returned an HTTP 404 error. Therefore, the article bases its metrics on the full paper and does not present the repository as a publicly available resource.<\/p>\n<aside class=\"td-article-note\"><strong>Critical measurement distinction:<\/strong> The approximately 1,260 tokens relate to the routing prompt\u2019s schema segment in prompt-all. The 207 tokens in the downstream RAG table relate to the retrieved context provided to the reader. They are neither the same stage nor the same cost unit.<\/aside>\n<h2 id=\"ektelesimo-tool-plan\">How an executable tool plan is generated<\/h2>\n<p>In the first step, a small LLM extracts the intent, domain concepts, parameter values, source constraint, and explicitly mentioned tools. If the user names a specific tool, that tool takes precedence. Otherwise, candidates are filtered based on the intent\u2019s domain and the required field groups. A source constraint can, for example, distinguish between experimental, calculated, or literature data.<\/p>\n<p>Field selection is then performed deterministically. This includes identifiers, concept-field matches, and field groups that serve as a recall safety net. The alias layer bridges the user\u2019s language with the canonical field names: a domain term can be mapped to the exact name understood by the API. Finally, the schema associates a unit, source type, and license with each relevant field.<\/p>\n<p>\u00abValid by design\u00bb does not mean that the business policy has been automatically implemented. It means that the fields come from the declared schema and the parameters from the set accepted by the endpoint. Authentication, authorization, consent, rate limits, data minimization, and approval gates must be enforced by the respective policy and execution layers. The <a href=\"https:\/\/twodots.gr\/crm-compliance-dedomena-pelaton-marketing\/\">Data protection in CRM without hindering marketing efforts<\/a> requires precisely this distinction between the usable area and permitted access.<\/p>\n<h2 id=\"benchmark-110-erotimata\">What Does the 110-Question Benchmark Really Measure?<\/h2>\n<p>The evaluation includes 110 materials science questions across nine categories: 36 single-tool property lookups, 12 chemical lookups, 12 provenance-sensitive queries, 10 multi-field queries, 10 literature searches, 8 cross-source comparisons, 8 unit-sensitive queries, 8 negative queries without a suitable tool, and 6 ambiguous single-entity queries.<\/p>\n<p>The frozen inventory covers 19 heterogeneous tools and 166 response fields: material property databases, PubChem, bibliographic APIs, a source of business disclosures, and web or Q&amp;A fallbacks. The downstream data consists of deterministic structured fixtures. This choice facilitates controlled comparison but underestimates the noise, clutter, and potential prompt injection found in a real-world corpus.<\/p>\n<p>The study goes beyond tool precision and F1 score. It measures parameter validity, executable-plan rate, retrieved-context tokens, end-to-end latency, final response accuracy, and whether the response includes provenance and license information. This is the right standard for a production agent: the router is not good because it produces an \u00abelegant\u00bb plan, but because it helps ensure that the final answer is correct, fast, and verifiable.<\/p>\n<h2 id=\"apotelesmata-ligotero-context\">The results: competitive accuracy with much less context<\/h2>\n<p>SchemaRouter achieved an answer accuracy of 0.71 with a 95% confidence interval of 0.63\u20130.79, a field recall of 0.86, 227 retrieved-context tokens, and an end-to-end latency of 5.5 seconds. Fetch-all had an accuracy of 0.70, but used 2,066 tokens and took 4.9 seconds. Prompt-all had an accuracy of 0.66, 207 retrieved-context tokens, and a latency of 15.1 seconds.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">The SchemaRouter's operating point<\/p>\n<p class=\"td-chart-subtitle\">The figures apply exclusively to the 110-question benchmark and do not constitute a guarantee for every agentic RAG deployment.<\/p>\n<\/div>\n<div class=\"td-chart-body\">\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">0,71<\/span><span class=\"td-metric-label\">Answer accuracy with overlapping confidence intervals<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">227<\/span><span class=\"td-metric-label\">retrieved-context tokens<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">0,93<\/span><span class=\"td-metric-label\">\"tool-exact rate\" in the main table<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">62%<\/span><span class=\"td-metric-label\">answers with a \"schema-only\" license citation<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>The clearest distinction lies in the model\u2019s efficiency and validity. SchemaRouter achieved a tool-exact rate of 0.93 and a parameter validity of 1.00 in the baseline planning metrics table. It used approximately nine times fewer retrieved-context tokens than fetch-all, without any measurable loss of accuracy in this specific setup. Compared to prompt-all, the end-to-end time was approximately 2.7 times shorter.<\/p>\n<aside class=\"td-article-note\"><strong>This is not a statistically proven victory in terms of accuracy:<\/strong> The confidence intervals of the top routers overlap. Furthermore, fetch-all was faster than SchemaRouter on this particular shared server. The paper\u2019s strongest argument is the combination of context efficiency, valid plans, and metadata fidelity\u2014not absolute first place on every metric.<\/aside>\n<h2 id=\"elachista-pedia-lathos-kpi\">Why the minimum number of fields is the wrong KPI<\/h2>\n<p>One of the most useful findings is that extreme field minimization harms the downstream performance. At the minimal-field operating point, accuracy dropped to 0.56 with 188 retrieved context tokens. When intent-relevant field groups were added as a safety net for recall, accuracy rose to 0.71, while the number of tokens increased to only 227.<\/p>\n<p>The lesson isn't \u00abbring more data.\u00bb It\u2019s that the correct optimization target combines answer quality and actual context cost. A field-count metric treats all fields as equivalent, whereas a short identifier and a long textual field have very different token footprints. The smaller plan may seem efficient but at the same time lack the element that allows the reader to answer.<\/p>\n<p>For product and data teams, dashboards should link routing metrics to response success, latency, token usage, invalid calls, and missing-field failures. Similarly, the <a href=\"https:\/\/twodots.gr\/multimodal-ai-agents-antilipsi-drasi\/\">Multimodal AI agents require verification after perception<\/a>: More inputs do not guarantee correct behavior, but neither does aggressively removing context guarantee that critical information will be preserved.<\/p>\n<h2 id=\"provenance-adeies-asfaleia\">Provenance, Licenses, and Data Security<\/h2>\n<p>The most significant difference at the answer level concerns traceability. SchemaRouter included schema-only license citations in 62% of non-negative answers. Both license mentions and source-type statements reached 94%. The baselines had a schema-only citation rate of approximately zero, because those specific license strings were not present in the text received by the reader.<\/p>\n<p>For marketing, e-commerce, and content operations, this principle extends beyond materials science. When an agent uses product feeds, DAM assets, customer data, or licensed market research, the source and terms of use should not be added as an afterthought in a footnote. They must travel alongside the data from the registry to the final output. This is an architectural interpretation, not a cross-domain result demonstrated by the benchmark.<\/p>\n<p>A smaller API response can also reduce the risk of excessive data exposure, but only if field selection is enforced on the server side and is linked to object-property authorization. A router that \u00abrequests\u00bb fewer fields does not replace access controls. The MCP specification also warns that tool annotations are hints and should not be considered, on their own, a reliable description of behavior.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-kicker\">Production gate for field-aware agents<\/p>\n<p class=\"td-decision-title\">The schema selects data; the policy determines whether it is allowed<\/p>\n<p>Associate each endpoint and response field with owner, sensitivity, permitted purpose, role, source, unit, license, and retention period. Execute the plan only after it has passed the authorization and data minimization checks, and keep a trace of the query, the selected schema version, the returned fields, and the final response.<\/p>\n<\/div>\n<h2 id=\"oria-schemarouter\">Where Does SchemaRouter Still Fall Short?<\/h2>\n<p>The remaining errors are mainly divided into incorrect tool selection for ambiguous entities and reader errors. In ambiguous single-entity queries, the concept matcher has difficulty determining the appropriate source without an explicit name. The alias layer requires human maintenance as a domain grows, so this approach does not eliminate the need for schema governance.<\/p>\n<p>There are significant validity limitations. The benchmark consists of 110 queries, is materials-science-centric, was created by a single annotator, and uses frozen structured fixtures instead of actual documents. The evaluation was conducted in a single run using a self-hosted Qwen-3.5 122B model, at a temperature of zero, with reasoning disabled. The same model family generated and evaluated the responses, while a cross-family spot review was performed on only 40 samples.<\/p>\n<p>The latency measurements were taken from a shared, heavily loaded server and are indicative. The results require multi-seed, multi-model, and cross-domain replication. Furthermore, the paper supports schema-size-independent routing prompts, but maintaining the graph, aliases, field groups, and policies remains a real organizational cost.<\/p>\n<h2 id=\"crm-ecommerce-marketing-agents\">What does this mean for CRM, e-commerce, and marketing agents?<\/h2>\n<p>A commerce agent can connect to catalogs, warehouses, pricing systems, orders, reviews, and ad platforms. The question \u00abWhich product has room for promotion?\u00bb requires specific fields and may require different access permissions than \u00abWhich campaign had the best ROAS?\u00bb If all descriptions and all data are always loaded, the context expands. If only the semantically closest tool is selected, the product cost or attribution window may be missing.<\/p>\n<p>A field-aware registry can specify which fields each intent requires, their units, source, freshness, and access constraints. For a CRM query, for example, the query may require the lifecycle stage and the most recent allowed contact, but not the full history of notes or information unrelated to the purpose. For product availability, it requires the SKU, stock status, and timestamp\u2014not the entire order item.<\/p>\n<p>The production deployment requires privacy and authorization enforcement, which are not covered in this paper. A fallback mechanism is also needed when an endpoint changes its schema, when the source data is stale, or when the plan does not meet a confidence threshold. The <a href=\"https:\/\/twodots.gr\/ai-agents-business-not-coworkers\/\">AI agents in business should be treated as software with limitations<\/a>, not as \u00abcolleagues\u00bb who arbitrarily fill in the blanks.<\/p>\n<h2 id=\"pilot-epta-vimata\">A seven-step pilot for field-aware routing<\/h2>\n<p>The pilot should start with a limited decision that has a measurable cost of error, not with the company\u2019s entire API inventory. A suitable first use case is a read-only query in a catalog, analytics, or support knowledge base, where there are known correct answers and a safe fallback.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">Seven Steps from Inventory to a Secure Shadow Run<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Step 1<\/span><strong>Lock in a business intent<\/strong>\n<p>Specify which question the agent must answer, which data it is prohibited from accessing, who the owner is, and what the cost of an error or delay is.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 2<\/span><strong>Map tools and endpoints<\/strong>\n<p>Document the method, required parameters, response fields, rate limits, versions, and failure modes only for the systems that require that specific intent.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 3<\/span><strong>Attach metadata by field<\/strong>\n<p>Specify the canonical name, aliases, unit, source type, freshness, license, sensitivity, and required role. Do not enter permissions as free text at the prompt.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 4<\/span><strong>Here are the recall-critical field groups<\/strong>\n<p>Identify identifiers and small groups of fields that are needed as a safety net, so that minimization does not result in under-fetching.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 5<\/span><strong>Build a Gold Query Suite<\/strong>\n<p>Add simple, multi-purpose, provenance-sensitive, ambiguous, and negative queries where the correct decision is to avoid making an inappropriate call.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 6<\/span><strong>Compare downstream baselines<\/strong>\n<p>Evaluate prompt-all, similarity routing, and controlled fetch-all based on answer accuracy, valid execution, context tokens, latency, metadata completeness, and data exposure.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Step 7<\/span><strong>Run shadow mode and versioned monitoring<\/strong>\n<p>Record the query, schema version, tool plan, returned fields, answer, and human correction before allowing write actions or autonomous impact on customers and the budget.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<p>The <a href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Business Automation &amp; AI by TWO DOTS<\/a> They start by mapping processes, data, owners, and checkpoints. For a SchemaRouter-style pilot, the value lies not in adding yet another abstraction layer, but in demonstrating that the plan reduces unnecessary context without losing critical fields or bypassing policies.<\/p>\n<h2 id=\"mikrotero-eparkes-synolo\">The Essence: The Smallest Sufficient and Verifiable Set<\/h2>\n<p>SchemaRouter does not prove that a schema graph is the best solution for every agentic RAG system. However, it does show that the selection of response fields is worth treating as a separate architectural problem. In this specific benchmark, recall-preserving routing maintained competitive accuracy with significantly less retrieved context than fetch-all, while providing richer information on provenance and permissions.<\/p>\n<p>For businesses, the ideal goal is neither \u00aball the data\u00bb nor \u00abas little as possible.\u00bb It is the smallest sufficient, executable, permissible, and verifiable set that enables a correct answer. This requires metrics that track the entire process: intent, tool selection, parameter validation, field projection, authorization, retrieval, answer, and human verification.<\/p>\n<p>If the schema graph remains versioned, aliases are maintained by domain owners, and permissions are enforced outside the model, this approach can serve as a useful control point. If treated solely as a token-reduction technique, it risks optimizing the wrong metric and shifting complexity to a registry that no one takes responsibility for maintaining.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-kicker\">From the inventory API to a controlled pilot<\/p>\n<p class=\"td-service-cta-title\">Design AI agents that request the right data<\/p>\n<p>TWO DOTS maps intents, tools, response fields, permissions, validation, and monitoring so that an AI workflow can reduce unnecessary context without losing critical information or control.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Design a safe AI pilot<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">Frequently Asked Questions (FAQs)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is SchemaRouter?<\/summary>\n<div class=\"td-faq-content\">\n<p>It is a research routing layer that organizes tools, endpoints, parameters, response fields, and metadata into a schema graph and generates executable tool plans.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What problem is it trying to solve?;<\/summary>\n<div class=\"td-faq-content\">\n<p>It attempts to limit both over-fetching\u2014which retrieves unnecessary tools or fields\u2014and under-fetching\u2014which omits information necessary for a correct response.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">How is it different from vector similarity routing?<\/summary>\n<div class=\"td-faq-content\">\n<p>Similarity can find a relevant tool, but it does not guarantee valid parameters or sufficient response fields. SchemaRouter applies deterministic field selection to the graph.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What benchmark results did it achieve?<\/summary>\n<div class=\"td-faq-content\">\n<p>On 110 queries, it achieved an answer accuracy of 0.71 with 227 retrieved-context tokens, a tool-exact rate of 0.93, and a parameter validity of 1.00 in the main table of the paper.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Is it really the most accurate router?<\/summary>\n<div class=\"td-faq-content\">\n<p>No. The confidence intervals of the top routers overlap. The paper primarily attributes the advantage to context efficiency, executable plans, and provenance or license metadata.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Why is the minimum number of fields a bad KPI?<\/summary>\n<div class=\"td-faq-content\">\n<p>Because at the minimal-field operating point, the accuracy dropped to 0.56. Adding recall-critical field groups restored it to 0.71, with a slight increase from 188 to 227 tokens.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Does SchemaRouter enforce permissions and privacy?<\/summary>\n<div class=\"td-faq-content\">\n<p>Not on its own. The schema can carry relevant metadata, but authorization, consent, data minimization, and approval gates must be enforced by the policy and execution layers.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is the first step in launching a business pilot?<\/summary>\n<div class=\"td-faq-content\">\n<p>Select a restricted read-only intent, record the necessary endpoints and fields, and create gold queries with correct answers, ambiguous cases, and safe negative examples.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">Sources<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.21375\" target=\"_blank\" rel=\"noopener\">arXiv \u2014 SchemaRouter: Field-Aware Tool Routing for Efficient Heterogeneous Agentic RAG<\/a><\/li>\n<li><a href=\"https:\/\/airc.nist.gov\/airmf-resources\/airmf\/5-sec-core\/\" target=\"_blank\" rel=\"noopener\">NIST AI RMF Core \u2014 Govern, Map, Measure, and Manage<\/a><\/li>\n<li><a href=\"https:\/\/airc.nist.gov\/airmf-resources\/playbook\/manage\/\" target=\"_blank\" rel=\"noopener\">NIST AI RMF Playbook \u2014 Provenance and Third-Party Source Management<\/a><\/li>\n<li><a href=\"https:\/\/wstg.owasp.org\/latest\/4-Web_Application_Security_Testing\/12-API_Testing\/03-Excessive_Data_Exposure\/\" target=\"_blank\" rel=\"noopener\">OWASP WSTG \u2014 Excessive Data Exposure in APIs<\/a><\/li>\n<li><a href=\"https:\/\/modelcontextprotocol.io\/specification\/2025-11-25\/schema\" target=\"_blank\" rel=\"noopener\">Model Context Protocol \u2014 Official Reference for Schema and Tool Annotations<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>SchemaRouter demonstrates how AI agents correctly select tools and fields, reduce context, and preserve provenance, units, and licenses.<\/p>","protected":false},"author":1,"featured_media":97919,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","footnotes":""},"categories":[199],"tags":[20455,6506,9265,20456,20424],"class_list":["post-97182","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-agentic-rag","tag-ai-agents","tag-data-governance","tag-schemarouter","tag-tool-routing"],"_links":{"self":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/97182","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/comments?post=97182"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/97182\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media\/97919"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media?parent=97182"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/categories?post=97182"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/tags?post=97182"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}