# Retrieval, agents & safety > Grounded generation, language-aware retrieval, agent behavior, policy evaluation, and safeguards across markets. [Multimodal AI capability](https://admas.net/capabilities/llm-solutions/index.md) ## The challenge Retrieval and safety systems inherit language asymmetries. The right source may not be found, policy meaning may drift, and safeguards can overblock or miss risk outside English. We test the full path from multilingual query to retrieved evidence, generated response, citation, policy decision, and user-facing refusal or recovery. The work connects information architecture, model behavior, language variation, and harm analysis in one product-level view. ## How Admas works 1. **Map the control path:** Trace queries, language detection, retrieval, reranking, generation, citations, moderation, and fallbacks. 2. **Challenge the system:** Build multilingual cases for retrieval gaps, code-switching, obfuscation, policy ambiguity, and cultural context. 3. **Harden and monitor:** Improve sources, routing, thresholds, prompts, safeguards, and ongoing market-specific evaluation. ## Typical outputs - Multilingual retrieval and safety assessment - Adversarial and representative challenge set - Control and policy recommendations - Monitoring metrics and release gates ## Frequently asked questions ### What is retrieval, agents & safety? Grounded generation, language-aware retrieval, agent behavior, policy evaluation, and safeguards across markets. In practice, the work is bounded by a defined product or model decision, named audiences and locales, representative inputs, and acceptance criteria that can be reviewed. ### When does a team need retrieval, agents & safety? Retrieval and safety systems inherit language asymmetries. The right source may not be found, policy meaning may drift, and safeguards can overblock or miss risk outside English. The useful starting point is the smallest representative flow that can expose the cause, impact, and ownership of the problem. ### What does a retrieval, agents & safety engagement include? Map the control path: Trace queries, language detection, retrieval, reranking, generation, citations, moderation, and fallbacks. Challenge the system: Build multilingual cases for retrieval gaps, code-switching, obfuscation, policy ambiguity, and cultural context. Harden and monitor: Improve sources, routing, thresholds, prompts, safeguards, and ongoing market-specific evaluation. ### What should we provide before retrieval, agents & safety starts? The most useful inputs are the product task and user journey, candidate models or system access, priority languages and communities, policies and risk thresholds, representative prompts, media, tools, and and expected outcomes. Admas can begin with a partial package, but missing context, rights, access, owners, or acceptance criteria will be made visible in the plan rather than treated as harmless assumptions. ### What does Admas deliver for retrieval, agents & safety? Typical outputs include multilingual retrieval and safety assessment, adversarial and representative challenge set, control and policy recommendations, and monitoring metrics and release gates. Deliverables are adapted to the team that must use them, with decisions, evidence, limitations, owners, and next actions made explicit. ### How is the quality of retrieval, agents & safety evaluated? Quality is measured against the real task and risk. Relevant evidence can include task success by language and scenario, human-rated meaning and usefulness, safety and policy performance, retrieval and citation fidelity, tool-call correctness, and regressions and disparities hidden by aggregate scores. Sampling, severity rules, reviewers, adjudication, and pass or fail thresholds should be agreed before the result is used as a release decision. ### Can AI replace the human work in retrieval, agents & safety? Models and automated checks can generate candidates, expand test sets, cluster failures, and accelerate analysis. Qualified humans define what good means, identify culturally or linguistically plausible failures, adjudicate close cases, and own consequential release judgments. The right allocation depends on consequence, content stability, available references, language coverage, reversibility, and the cost of a plausible but wrong result. ### How much does retrieval, agents & safety cost? The estimate changes with number of languages, modalities, systems, and scenarios, risk level, dataset creation needs, evaluator specialization, sampling and adjudication depth, and experiment and reporting cadence. Pricing should distinguish setup and discovery, repeatable units, specialist or engineering time, independent review, management, and external costs. A low unit price is not comparable if it excludes the QA cycle or shifts rework back to the buyer. ## Related multimodal ai services - [Multimodal AI evaluation](https://admas.net/capabilities/llm-solutions/multilingual-evaluation/index.md): Task-grounded benchmarks and human evaluation for multilingual text, audio, image, and video model behavior. - [Model localization & adaptation](https://admas.net/capabilities/llm-solutions/model-adaptation/index.md): Data and adaptation strategies that improve model behavior for specific languages, cultures, modalities, and product tasks. - [Multilingual agent engineering](https://admas.net/capabilities/llm-solutions/multilingual-agent-engineering/index.md): Agent architectures that preserve locale, language, meaning, and policy across prompts, memory, retrieval, tools, and actions. - [Multimodal safety evaluation](https://admas.net/capabilities/llm-solutions/multimodal-safety-evaluation/index.md): Language- and culture-aware safety evaluation across text, speech, images, video, retrieval, and model-mediated actions. ## Start a project - [Build a project brief](https://admas.net/start-a-project/index.md?focus=llm-solutions): Tell Admas what you are building, which modalities and languages matter, and where progress is blocked.