File search tool

The file search tool extends the models' capabilities by enabling hybrid search through the user's files when generating the response. You can enable the tool in the Responses API. For the prices, see Yandex AI Studio pricing policy.

To allow the model to search for information through the files, specify the "file_search" tool and the Vector Store search index ID. The search index contains the information on the files that will be used for the search. You can use the max_num_results parameter to limit the number of results you get.

response = client.responses.create(
    model=f"gpt://{YANDEX_CLOUD_FOLDER}/{YANDEX_CLOUD_MODEL}",
    # SpeechKit Voice Profiler is a non-existent product (at least for now) which was described in the documents.
    instructions="You are helping the user with the documentation. When asked about SpeechKit Voice Profiler, use file search.",
    input="Tell me about SpeechKit Voice Profiler",
    tools=[{
        "type": "file_search",
        "vector_store_ids": [<search_index_ID>],
        "max_num_results": 3
    }]
)

AnnotationsAnnotations

The response with the file search results is inside the annotations object. Use it to track and show the data sources to the user. The filename field of the annotations object contains the name of the file the information was found in, and the file_id field gives its ID. The object's other fields are populated with default values for compatibility with OpenAI:

"annotations": [
  {
    "file_id": "...",
    "filename": "Yandex SpeechKit Voice Profiler.docx",
    "index": 0,
    "type": "file_citation"
  }
 ...
]

Possible use casesPossible use cases

The file search tool can prove useful in various scenarios:

• Corporate assistant: Advice based on internal documents and corporate regulations.
• Product support: Searching through technical documentation, SDKs, and FAQs with links to relevant sections.
• Legal consulting: Advice based on internal legal documents, with quotes and sources included in the response.
• New staff onboarding: Answers to trainee's questions on internal policies and project architecture.
• Sales: Searching for successful deals in the presentations and bids directory.

See alsoSee also

Creating a text agent with file search