Foundry IQ Welcomes Microsoft-Retriever-1 and Graphs

Microsoft adds knowledge graphs and the Microsoft-Retriever-1 model to Foundry IQ, while a managed service is announced for public preview in December 2026. These additions aim to facilitate the creation of agents capable of accessing, understanding, and reasoning about enterprise data. The anticipated benefits come with caveats regarding the quality of extracted relationships, costs, and technical dependency.
Accuracy of Graphs and Cost Predictability
Business vocabulary, acronyms, or variations in names can mislead automatic extraction and lead agents to incorrect conclusions, warns Ashish Chaturvedi, head of research at HFS Research. He considers the accuracy of entities and their relationships as the primary point of vigilance for graphs. He therefore recommends verifying their alignment with actual business links and updating them in line with document changes.
Another concern from the analyst is that generating a response from a graph may be slower and more costly than traditional search. He believes this technique is more suited for complex requests than for routine inquiries. With Microsoft-Retriever-1, successive searches could also increase delays and expenses, as each iteration adds search and processing operations. The benefit would thus depend on the model's ability to stop once the necessary elements are gathered.
Ashish Chaturvedi advises testing on company-specific data: in his view, Microsoft has not provided detailed performance comparisons. He also warns that the model's choices may be less straightforward to examine than an explicitly programmed search mechanism. The analyst particularly notes that the default reliance on Microsoft’s retrieval technology in new Foundry projects could heighten companies' dependency on it.
Stephanie Walter, head of the AI stack practice at HyperFrame Research, cautions that continuous consumption by agents could make billing less predictable. She recommends budget caps and rules determining which projects can use the service.
Graphs Connect Information Across Documents
Announced during Microsoft Ignite, the graph feature is already available in public preview. Developers can attach a configuration that includes their ontology to an existing knowledge source or import a pre-existing graph. Microsoft states that Foundry IQ relies on this ontology to link documents, contacts, and information while maintaining the provenance of each connection.
According to Microsoft, these links between entities and facts, as well as the categorization of content by topics, should help agents handle requests that mobilize multiple sources.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Ashish Chaturvedi illustrates the difference with a regulation that could affect certain products. A traditional search finds, in his view, documents close to the request without traversing their links: it can return the rule and texts that mention it but leaves the developer to connect regulation, supplier, and product. With a graph, he explains, the agent can traverse these relationships and cite the steps, requiring less manual assembly and specific programming.
The analyst believes that a more comprehensive document context could enhance decision-making and the reliability of executing complex processes. He cites potential applications such as regulatory compliance, supply chain risks, contract review, and due diligence for mergers or acquisitions.
The Model Chooses How to Continue Searches
With its specialized model, Microsoft aims to limit searches and the reliance on large language models to organize the process. The reduction of latency and the number of tokens consumed is a stated goal, not a quantified result. According to the company, Microsoft-Retriever-1 interprets the request, reformulates it with terms present in internal documents, and decides if further searches are required.
Microsoft clarifies that activation is possible on existing indexes and knowledge bases without additional provisioning or new document indexing. Stephanie Walter believes the model could reduce the code needed when agents query multiple internal sources. She attributes decisions about where to search, the next request, and when evidence is sufficient to the model, rather than leaving these choices to manually written logic.
For the analyst, entrusting retrieval to a specialized model rather than systematically calling on large generalist models could lower costs and improve responsiveness. These potential savings would make the deployment of agents financially feasible across a broader range of processes.
The Managed Service Will Equip Every New Project
Scheduled for public preview in December 2026, the system will automatically create a managed Foundry IQ resource for every new Foundry project. Microsoft announces that it will handle the necessary infrastructure for retrieval with serverless billing tied to consumption. It would no longer be necessary to separately create and manage an Azure AI Search resource to build knowledge bases and connect agents to internal information.
Ashish Chaturvedi reminds us that this work typically involves setting up a search service, connecting sources, and configuring indexes before actual development begins. He believes that the automatic preparation of infrastructure would reduce these tasks and allow for immediate prototyping, billed according to usage.
For Ashish Chaturvedi, starting projects with a common service could also standardize access to information and its governance. IT departments would have fewer distinct internal configurations to protect and maintain. He finally estimates that experimentation could be cheaper since unused projects would not have to bear ongoing infrastructure costs.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.