WebAI
WebAIExperiment
Knowledge preview

Experiment

A laboratory of ideas and mechanisms.

WebAI LabSimulationRAG
UNDERSTAND THE TOPIC

Transparent learning simulations show how a mechanism works without pretending to be a connected service. The first module below illustrates document selection before an answer is assembled.

What you will discover

  • Choose a question and inspect the selected source snippets.
  • Observe the limits of answers built from retrieved documents.
  • Distinguish a simulation from a live model.
SIMULATION · NO CONNECTED AI

How a RAG system uses its documents

Deterministic example

Choose a question. This module shows preselected snippets and a prewritten answer derived from them. No real vector retrieval or model generation takes place.

1. Choose a question

2. Selected snippets

Choose a question to display an example.

3. Example answer

—

Teaching limitation: in a real RAG system, retrieval quality, synthesis fidelity and source relevance need to be evaluated separately.

No live news collection or AI model connection is enabled in this preview.

DEEPEN YOUR KNOWLEDGE · RESEARCH DOSSIERS

In-depth explanations, concrete cases and their evidence

Every dossier connects mechanisms, cases, limitations and original documents. Independent scientific review remains pending; original sources are provided for verification.

5studies to explore
01 Recherche

Rigorous evaluation of RAG systems

Identify errors at each stage of retrieval-augmented generation.

Index evidence, not just text

A RAG system selects candidate passages and supplies them to a model to draft an answer. Document identity, edition, date and rights should remain attached to indexed fragments; otherwise citations lose context.

Separate retrieval recall from faithfulness

A system can retrieve relevant passages yet write an unfaithful answer, or generate plausible prose from irrelevant evidence. Evaluate retrieval, reranking, faithfulness, citation accuracy and abstention when evidence is missing separately.

View the verification workflow
  1. Prepare reference questions
  2. Measure retrieval separately
  3. Evaluate answer faithfulness
  4. Record corpus and model versions
Explore explanations, steps and primary sources Read the primary source
02 Agents

WebMCP: JavaScript tools exposed by a page

Study the experimental API and the exact status of the community draft.

What WebMCP proposes

WebMCP proposes a JavaScript interface for pages to expose functions as callable tools for agents. Natural language descriptions and structured input schemas help discover features normally activated by users.

Status as of 13 November 2026

The Web Machine Learning Community Group report remains a community draft rather than a W3C Recommendation. The imperative API is documented; a declarative approach built around HTML forms is explored separately and must not be described as universally implemented.

View the verification workflow
  1. Read the community draft
  2. Model a harmless action
  3. Check actual browser support
  4. Review permission boundaries
Explore explanations, steps and primary sources Read the primary source
03 Exécution

WebGPU for computation and Web models

Separate GPU computation from model-level AI logic.

A GPU computation API

WebGPU lets Web applications render graphics and perform certain computations on GPUs. As of 13 November 2026 the consulted W3C publication is a Candidate Recommendation Draft; availability is not identical across all devices.

From GPU to inference

AI runtimes may use WebGPU for particular operations. WebGPU alone contains neither a model nor a dataset; it provides computation primitives.

View the verification workflow
  1. Inspect GPU support
  2. Measure performance and memory
  3. Implement a fallback
  4. Document browser support
Explore explanations, steps and primary sources Read the primary source
04 Modèles

Built-in browser AI: capabilities and limits

Assess translation, summarization and generation APIs offered by some browsers.

Browser-managed models

Some Chrome features expose browser-managed models for summarizing, translating, language detection or text production. Availability depends on browser version, hardware and conditions; websites cannot assume universal support.

Local processing and its nuances

Local inference can reduce some data transfers and support offline uses once required resources are available. Initial model downloads, memory, battery life and browser policies still impose constraints.

View the verification workflow
  1. Detect supported browser features
  2. Provide a usable fallback
  3. Compare quality and resource demands
  4. Document availability limitations
Explore explanations, steps and primary sources Read the primary source
05 Données

Structured data, JSON-LD and discoverability

Understand relationships between semantics, metadata and AI systems.

Linked data on the Web

JSON-LD represents resources and relations in JSON with a context defining identifiers. Linked data is neither an agent instruction nor a promise that a model will reproduce the content.

Schema.org and search

Page descriptions can clarify resource type, author, date or relationships. Google requires structured data to match visible content; artificial or misleading markup does not necessarily improve search eligibility.

View the verification workflow
  1. Choose stable identifiers
  2. Define JSON-LD context
  3. Check visible content consistency
  4. Test the graph
Explore explanations, steps and primary sources Read the primary source

Experimental proposals, published standards and educational interpretations are distinguished in the dossiers. Research snapshot: 13 November 2026.

Explore the full encyclopedia ↗