Microsoft Founders Hub gives startups Azure credits, access to experts, and other product benefits at any stage, including the idea stage.
2
AI templates package complex generative AI applications so startups can run, customize, and deploy them through GitHub Codespaces and Azure in minutes.
3
Reliable retrieval-augmented generation needs good retrieval, with hybrid text and vector search, evaluation, and often semantic ranking rather than vector search alone.
Summary
This workshop introduces Microsoft Founders Hub and then uses AI templates to build generative AI applications. Gabriela and Aishwarya explain that Founders Hub offers Azure credits, product benefits, and access to Microsoft experts. The workshop pass adds a temporary Azure subscription and an Azure OpenAI proxy so attendees can work without waiting for Azure OpenAI approval. Pamela Fox demonstrates GitHub Codespaces, a simple chat application, and deployment to Azure Container Apps with infrastructure-as-code files. The later examples use retrieval-augmented generation for product data in PostgreSQL and for unstructured documents such as PDFs. The document workflow uses Azure Document Intelligence, token-based chunking, embeddings, and Azure AI Search. The speakers also discuss query rewriting, SQL filters, citations, and automatic index updates. Pamela is direct about retrieval quality: vector search always returns something, even when the result is poor, so applications should evaluate retrieval and use hybrid search with semantic ranking.
Founders Hub is available to startups before they have funding or incorporation
Gabriela says Founders Hub is open to startups at any stage. Someone with only an idea can apply, as can an incorporated company with or without funding. The platform includes Azure credits, product benefits, and guidance. Its expert matching feature can connect founders with people who answer technical questions about infrastructure and Kubernetes, as well as questions about go-to-market plans and subscription pricing. Gabriela says the signup takes less than five minutes. Aishwarya adds that the platform tracks credits, products, and benefits, including third-party credits and productivity tools.
Microsoft's startup support includes advice and product collaboration
Aishwarya says the startup program is not only a credit allocation. Microsoft product teams work with startups during private previews, collect feedback about missing capabilities, and pass that feedback to the product team. Her AI advisory team works with startups one-on-one, while Founders Hub can match companies with experts across infrastructure, Kubernetes, pricing, and go-to-market questions. She says this support matters because startups often lack enough internal resources. The program follows companies from the idea stage through building and scaling.
AI templates target the long setup time of complex applications
Aishwarya describes the problem behind AI templates. Startups want to experiment with different tools and models quickly, but a first end-to-end application can take weeks when developers work through separate documentation and quick-start examples. The templates address this by providing complex examples that can run within minutes. One example is retrieval-augmented generation with AI Search. The repositories are open source, so users can customize the applications and file issues on GitHub. A cloud advocate team maintains the templates and accepts requests for additional examples.
GitHub Codespaces gives workshop users a shared development environment
Pamela opens the template in GitHub Codespaces, which runs VS Code in the browser with the project already loaded. She explains that Codespaces helps avoid the familiar situation where software works on one developer's machine but not another's. The environment can be customized with the repository, so different users start from the same setup. She first runs the application inside the Codespace and connects it to the Azure OpenAI proxy. The templates use asynchronous Python frameworks such as Quart or FastAPI because applications that call language models need concurrency.
The templates support local testing before Azure deployment
The first application is a simple chat app. Pamela creates an environment file with the proxy endpoint, API key, and deployment name, then starts the Python backend inside the Codespace. She tests it with short prompts and explains that the application streams model output back to the browser. Streaming lets the user see the first token as soon as it arrives instead of waiting for the complete response. After local testing, the deployment flow uses Azure CLI login, creates an Azure Container Apps environment, sets deployment variables, and runs 'azd up'.
RAG combines a user's question, retrieved data, and a language model
Pamela defines retrieval-augmented generation as a way to give a language model information from a particular domain. The application uses the user's question to search a database or search engine, then sends both the original question and the retrieved results to the model. The model is instructed to answer from those sources. One template applies this pattern to product rows in PostgreSQL. Another handles unstructured documents such as PDFs, Word files, Excel files, and HTML. The document version returns citations that point users to the source document and page number.
Hybrid retrieval is safer than relying on vector search alone
For the PostgreSQL example, Pamela combines vector search with PostgreSQL full-text search. She says vector search alone can return a result even when the result is far from the query, which can distract the language model. A hybrid search combines text and vector results, while Azure AI Search's semantic ranker can reorder them against the original user query. Pamela advises evaluating these choices rather than assuming vectors will improve an application. In one evaluation she cites, text-only search had a groundedness rating of 4.87, while adding vector search improved it by only 0.2.
Document RAG needs an ingestion pipeline before users can search files
The document template first processes files into searchable content. Azure Document Intelligence extracts text from PDFs and other supported formats, including text found in images through OCR. The application then splits the extracted content into chunks of about 500 tokens, creates embeddings with Azure OpenAI, and stores the chunks in Azure AI Search. Pamela calls this the most complicated architecture because of the ingestion stage. The same approach can work with employee handbooks, blog sites, meeting transcripts, sales manuals, and other document collections. Integrated vectorization or an Azure Function trigger can refresh the index when source files change.
"We always want to use an async framework when we're building applications that make calls to LLMs because we want better concurrency."33:23
Who should watch
You are building a startup and need a working generative AI prototype without spending weeks on environment setup and cloud configuration.
Your application needs to answer questions over product records, PDFs, internal documents, or other private data, and you want a concrete RAG implementation to start from.
You are using vector search because it is popular and need a practical warning about hybrid retrieval, semantic ranking, and evaluation.
