Do knowledge files in Azure AI Foundry require File Search tool (especially with GPT-5 vs GPT-4)?

1. Knowledge files in Azure AI Foundry agents are backed by the File search tool and its vector store. For agents to search and use uploaded documents at runtime, the File search capability must be in place, because it is what parses, chunks, embeds, stores, and retrieves content from those files. File search is the mechanism that enables agents to “answer questions from proprietary or confidential content” and “search through documents you upload directly (PDFs, Word docs, code files)” without external infrastructure. Without this retrieval mechanism, models only see what is in the prompt and cannot dynamically query the uploaded files.
2. File search performs hybrid retrieval (keyword + semantic) over a vector store with default chunking (800-token chunks, 400-token overlap, up to 20 chunks in context) and query rewriting/reranking before injecting results into the model context. This is a tool-based retrieval pattern, not an implicit feature of GPT-4 or GPT-5 themselves. Any apparent behavior where GPT-4 “works” without explicitly enabling File search is due to how the agent configuration or portal wiring is set up (for example, a default or implicit file-search-backed knowledge base), not because GPT-4 can natively access knowledge files without a retrieval tool. The model alone does not index or search uploaded files.
3. GPT-5 is not described as having a different retrieval mechanism in the provided material; the retrieval behavior is defined by the File search tool and vector store, not by the model version. The tool is responsible for:
   • Query rewriting
   • Breaking complex queries into parallel searches
   • Hybrid search over vector stores
   • Reranking and selecting chunks to send to the model
   This means both GPT-4 and GPT-5 rely on the same tool-based retrieval pattern when using Foundry agents with knowledge files. Any stricter dependence observed with GPT-5 is an effect of configuration (for example, needing the File search tool attached) rather than a model-specific implicit retrieval capability.
4. File search uses a vector store with explicit limits and behavior:
   • Each vector store can hold up to 10,000 files.
   • Only one vector store can be attached to an agent and one to a conversation.
   • Supported file types and encodings are constrained (for example, .pdf, .docx, .json, .md, .txt, .py, etc., with UTF-8/UTF-16/ASCII for text MIME types).
   If files are outside these constraints, or if the vector store is not correctly attached, retrieval will not occur even if the files are uploaded.
5. The context provided does not describe a built-in tracing or inspection feature that exposes detailed File search tool calls (such as which chunks were retrieved, or why a tool was not triggered) for Foundry agents. It only describes the internal behavior (query rewriting, hybrid search, reranking) and the limits. Any deeper tracing (for example, logging tool invocations and retrieved chunks) would need to rely on capabilities outside the provided documentation.

In practice, to ensure knowledge files are used at runtime with both GPT-4 and GPT-5 in Azure AI Foundry agents:

• Ensure the File search tool (vector store) is configured and attached to the agent.
• Ensure files are in supported formats/encodings and within vector store limits.
• Use File search when the scenario matches: searching uploaded documents, answering from proprietary content, or processing large files without external search infrastructure.

References:

• File search tool for agents (python)
• File search tool for agents (csharp)
• Azure OpenAI Assistants file search tool (Preview) (classic)
• File search tool for agents (rest)