Create an OpenAI vector store, ingest files, and use the file_search tool on the Responses API to return answers grounded in your own documents with file citations

Documented steps

1. Create a vector store (POST the vector stores endpoint, e.g. client.vector_stores.create(name='knowledge_base')) and keep the returned vector_store id.
2. Upload each source document via the Files API with purpose='assistants'. Supported types include PDF, DOCX, DOC, PPTX, TXT, Markdown, HTML, JSON, CSS, LaTeX, and common code files.
3. For text/* MIME types, ensure the encoding is utf-8, utf-16, or ascii, or ingestion can fail.
4. Attach each uploaded file to the store: client.vector_stores.files.create(vector_store_id=..., file_id=...).
5. Poll ingestion with client.vector_stores.files.list(vector_store_id=...) and wait until every file's status is 'completed' before querying.
6. POST https://api.openai.com/v1/responses with tools:[{"type":"file_search","vector_store_ids":["<vector_store_id>"]}].
7. Optionally cap retrieval with max_num_results to cut token usage and latency.
8. Optionally filter by metadata, e.g. filters:{type:'in', key:'category', value:['blog','announcement']}.
9. Raw search results are NOT returned by default. Pass include:["file_search_call.results"] on the request to inspect matched chunks.
10. Read the 'file_search_call' output item for search metadata and the message item's annotations for file citations containing file_id, filename, and character indices.
11. Track ongoing storage cost per GB per day and delete unused vector stores; verify current rates on the pricing page.