Last episode we built semantic search from scratch. This time we let the AI provider host the vector store and handle the search. Upload PDFs, full policy docs, or markdown files, and the agent queries them through the SDK's FileSearch tool with no search logic on our side.

Why bother with complicated regular expressions when you can point and click your way to finding the files and folders you need?

AI Agent Framework, the Pydantic way. Contribute to pydantic/pydantic-ai development by creating an account on GitHub.

Microsoft is working on new Microsoft 365 Companion apps for Windows 11, and one of them is a so-called File Search Companion app that promises to "bring files right to your taskbar."

Last episode we built semantic search from scratch. This time we let the AI provider host the vector store and handle the search. Upload PDFs, full policy docs, or markdown files, and the agent queries them through the SDK's FileSearch tool with no search logic on our side.

Open-source semantic document search (RAG) engine with FastAPI and instant self-hosted deployment - flamehaven01/Flamehaven-Filesearch

Image:MicrosoftWindows11で2つ、新しいAI機能が使えるようになるそうです。2025年4月8日、MicrosoftがBlogで発表しました。で、どっちも便利そうなんですよ。1.ファイルを「内容で検索」できるAIファイル検索機能（Filesearch）は、AIとのチャットからファイル検索できるというもの。ファイル名だけでなく、“ファイル内の内容”から目的のファイルを探し出せるの

OpenAI’s file search does have two parameters to configure the `file_search` tool response size the AI model receives back. First, like you desire, one that promises to control the number of chunks returned (not files) - and there’s a possibility it could be working now if you want to try it: 'file_search' Tool Broken (Assistants API) > I also verified that despite specifying `"max_num_results": 6` and it appearing in the run steps for assistant, that 19 chunks were returned, the assistants endpoint blowing through the gpt-4-turbo budget I wanted to spend to replicate any issues. The other parameter, that would serve better to reject irrelevant results - a semantic comparison threshold: { "model": "gpt-4.1", "tools": [{ "type": "file_search", "vector_store_ids": ["vs_1234567890"], "ranking_options": {"score_threshold": 0.40}, }],... Increasing the number ranged 0.0-1.0 increases the similarity score required for ranked chunks to appear. OpenAI has no direct documentation of using `score_threshold` as plainly as my sample above: you have to expand the API reference several times to see API parameter inputs under “tools” to see it exists for the Responses API endpoint.

OpenAI’s file search does have two parameters to configure the `file_search` tool response size the AI model receives back. First, like you desire, one that promises to control the number of chunks returned (not files) - and there’s a possibility it could be working now if you want to try it: 'file_search' Tool Broken (Assistants API) > I also verified that despite specifying `"max_num_results": 6` and it appearing in the run steps for assistant, that 19 chunks were returned, the assistants endpoint blowing through the gpt-4-turbo budget I wanted to spend to replicate any issues. The other parameter, that would serve better to reject irrelevant results - a semantic comparison threshold: { "model": "gpt-4.1", "tools": [{ "type": "file_search", "vector_store_ids": ["vs_1234567890"], "ranking_options": {"score_threshold": 0.40}, }],... Increasing the number ranged 0.0-1.0 increases the similarity score required for ranked chunks to appear. OpenAI has no direct documentation of using `score_threshold` as plainly as my sample above: you have to expand the API reference several times to see API parameter inputs under “tools” to see it exists for the Responses API endpoint.

OpenAI’s file search does have two parameters to configure the `file_search` tool response size the AI model receives back. First, like you desire, one that promises to control the number of chunks returned (not files) - and there’s a possibility it could be working now if you want to try it: 'file_search' Tool Broken (Assistants API) > I also verified that despite specifying `"max_num_results": 6` and it appearing in the run steps for assistant, that 19 chunks were returned, the assistants endpoint blowing through the gpt-4-turbo budget I wanted to spend to replicate any issues. The other parameter, that would serve better to reject irrelevant results - a semantic comparison threshold: { "model": "gpt-4.1", "tools": [{ "type": "file_search", "vector_store_ids": ["vs_1234567890"], "ranking_options": {"score_threshold": 0.40}, }],... Increasing the number ranged 0.0-1.0 increases the similarity score required for ranked chunks to appear. OpenAI has no direct documentation of using `score_threshold` as plainly as my sample above: you have to expand the API reference several times to see API parameter inputs under “tools” to see it exists for the Responses API endpoint.