chore: import upstream snapshot with attribution

python/packages/azure-contentunderstanding/agent_framework_azure_contentunderstanding/__init__.py

@@ -0,0 +1,28 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Azure Content Understanding integration for Microsoft Agent Framework.
+
+Provides a context provider that analyzes file attachments (documents, images,
+audio, video) using Azure Content Understanding and injects structured results
+into the LLM context.
+"""
+
+import importlib.metadata
+
+from ._context_provider import ContentUnderstandingContextProvider
+from ._file_search import FileSearchBackend
+from ._models import AnalysisSection, DocumentStatus, FileSearchConfig
+
+try:
+    __version__ = importlib.metadata.version(__name__)
+except importlib.metadata.PackageNotFoundError:
+    __version__ = "0.0.0"
+
+__all__ = [
+    "AnalysisSection",
+    "ContentUnderstandingContextProvider",
+    "DocumentStatus",
+    "FileSearchBackend",
+    "FileSearchConfig",
+    "__version__",
+]

python/packages/azure-contentunderstanding/agent_framework_azure_contentunderstanding/_context_provider.py

@@ -0,0 +1,905 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Azure Content Understanding context provider using ContextProvider.
+
+This module provides ``ContentUnderstandingContextProvider``, built on the
+:class:`ContextProvider` hooks pattern.  It automatically detects file
+attachments, analyzes them via the Azure Content Understanding API, and
+injects structured results into the LLM context.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import re
+import sys
+import time
+from datetime import datetime, timezone
+from typing import TYPE_CHECKING, Any, ClassVar, TypedDict
+
+from agent_framework import (
+    AGENT_FRAMEWORK_USER_AGENT,
+    Content,
+    ContextProvider,
+    FunctionTool,
+    Message,
+    SessionContext,
+)
+from agent_framework._sessions import AgentSession
+from agent_framework._settings import load_settings
+from azure.ai.contentunderstanding import to_llm_input
+from azure.ai.contentunderstanding.aio import ContentUnderstandingClient
+from azure.ai.contentunderstanding.models import AnalysisInput, AnalysisResult
+from azure.core.credentials import AzureKeyCredential
+from azure.core.credentials_async import AsyncTokenCredential
+
+if TYPE_CHECKING:
+    from agent_framework._agents import SupportsAgentRun
+
+from ._detection import (
+    detect_and_strip_files,
+)
+from ._models import AnalysisSection, DocumentEntry, DocumentStatus, FileSearchConfig
+
+if sys.version_info >= (3, 11):
+    from typing import Self  # pragma: no cover
+else:
+    from typing_extensions import Self  # pragma: no cover
+
+logger = logging.getLogger("agent_framework.azure_contentunderstanding")
+
+AzureCredentialTypes = AzureKeyCredential | AsyncTokenCredential
+
+# Mapping from media type prefix to the appropriate prebuilt CU analyzer.
+# Used when analyzer_id is None (auto-detect mode).
+MEDIA_TYPE_ANALYZER_MAP: dict[str, str] = {
+    "audio/": "prebuilt-audioSearch",
+    "video/": "prebuilt-videoSearch",
+}
+DEFAULT_ANALYZER: str = "prebuilt-documentSearch"
+
+# Matches the leading YAML front-matter block emitted by ``to_llm_input``.
+# A rendered text with no markdown body (e.g. when the CU result has empty
+# ``markdown`` and no fields) is recognised by an empty tail after this match.
+# Accept both LF and CRLF line endings so body detection works cross-platform.
+_FRONT_MATTER_RE: re.Pattern[str] = re.compile(r"\A---\r?\n.*?\r?\n---(?:\r?\n|\Z)", flags=re.DOTALL)
+
+
+def _has_renderable_body(text: str) -> bool:
+    """Return True when ``text`` has any non-whitespace content beyond YAML front matter.
+
+    Used to skip ``file_search`` uploads when CU produced a result with no
+    markdown content — uploading a front-matter-only stub would pollute the
+    vector store without giving the LLM anything searchable.
+    """
+    if not text:
+        return False
+    match = _FRONT_MATTER_RE.match(text)
+    if match is None:
+        return bool(text.strip())
+    return bool(text[match.end() :].strip())
+
+
+class ContentUnderstandingSettings(TypedDict, total=False):
+    """Settings for ContentUnderstandingContextProvider with auto-loading from environment.
+
+    Settings are resolved in this order: explicit keyword arguments, values from an
+    explicitly provided .env file, then environment variables with the prefix
+    ``AZURE_CONTENTUNDERSTANDING_``.
+
+    Keys:
+        endpoint: Azure AI Foundry endpoint URL.
+            Can be set via environment variable ``AZURE_CONTENTUNDERSTANDING_ENDPOINT``.
+    """
+
+    endpoint: str | None
+
+
+class ContentUnderstandingContextProvider(ContextProvider):
+    """Context provider that analyzes file attachments using Azure Content Understanding.
+
+    Automatically detects supported file attachments in the agent's input,
+    analyzes them via CU, and injects the structured results (markdown, fields)
+    into the LLM context. Supports multiple documents per session with background
+    processing for long-running analyses. Optionally integrates with a vector
+    store backend for ``file_search``-based RAG retrieval on LLM clients that
+    support it.
+
+    Args:
+        endpoint: Azure AI Foundry endpoint URL
+            (e.g., ``"https://<your-foundry-resource>.services.ai.azure.com/"``).
+            Can also be set via environment variable
+            ``AZURE_CONTENTUNDERSTANDING_ENDPOINT``.
+        credential: An ``AzureKeyCredential`` for API key auth or an
+            ``AsyncTokenCredential`` (e.g., ``DefaultAzureCredential``) for
+            Microsoft Entra ID auth.
+        analyzer_id: A prebuilt or custom CU analyzer ID. When ``None``
+            (default), a prebuilt analyzer is chosen automatically based on
+            the file's media type: ``prebuilt-documentSearch`` for documents
+            and images, ``prebuilt-audioSearch`` for audio, and
+            ``prebuilt-videoSearch`` for video.
+            Analyzer reference: https://learn.microsoft.com/azure/ai-services/content-understanding/concepts/analyzer-reference
+            Prebuilt analyzers: https://learn.microsoft.com/azure/ai-services/content-understanding/concepts/prebuilt-analyzers
+        max_wait: Max seconds to wait for analysis before deferring to background.
+            ``None`` waits until complete.
+        output_sections: Which CU output sections to pass to LLM.
+            Defaults to ``["markdown", "fields"]``.
+        file_search: Optional configuration for uploading CU-extracted markdown to
+            a vector store for token-efficient RAG retrieval. When provided, full
+            content injection is replaced by ``file_search`` tool registration.
+            The ``FileSearchConfig`` abstraction is backend-agnostic — use
+            ``FileSearchConfig.from_openai()`` or ``FileSearchConfig.from_foundry()``
+            for supported providers, or supply a custom ``FileSearchBackend``
+            implementation for other vector store services.
+        source_id: Unique identifier for this provider instance, used for message
+            attribution and tool registration. Defaults to ``"azure_contentunderstanding"``.
+        env_file_path: Path to a ``.env`` file for loading settings.
+        env_file_encoding: Encoding of the ``.env`` file.
+
+    Per-file ``additional_properties`` on ``Content`` objects:
+        The provider reads the following keys from
+        ``Content.additional_properties`` (passed via ``Content.from_data()``
+        or ``Content.from_uri()``):
+
+        ``filename`` (str):
+            The document key used for tracking, status, and LLM references.
+            Without a filename, a UUID-based key is generated.
+            Must be unique within a session — uploading a file with a
+            duplicate filename will be rejected and the file will not be
+            analyzed.
+
+        ``analyzer_id`` (str):
+            Per-file analyzer override. Takes priority over the provider-level
+            ``analyzer_id``. Useful for mixing analyzers in the same turn
+            (e.g., ``prebuilt-invoice`` for invoices alongside
+            ``prebuilt-documentSearch`` for general documents).
+
+        ``content_range`` (str):
+            Subset of the input to analyze. For documents, use 1-based page
+            numbers (e.g., ``"1-3"`` for pages 1-3, ``"1,3,5-"`` for pages
+            1, 3, and 5 onward). For audio/video, use milliseconds
+            (e.g., ``"0-60000"`` for the first 60 seconds).
+
+        Example::
+
+            Content.from_data(
+                pdf_bytes,
+                "application/pdf",
+                additional_properties={
+                    "filename": "invoice.pdf",
+                    "analyzer_id": "prebuilt-invoice",
+                    "content_range": "1-3",
+                },
+            )
+    """
+
+    DEFAULT_SOURCE_ID: ClassVar[str] = "azure_contentunderstanding"
+    DEFAULT_MAX_WAIT_SECONDS: ClassVar[float] = 5.0
+
+    def __init__(
+        self,
+        *,
+        endpoint: str | None = None,
+        credential: AzureCredentialTypes | None = None,
+        client: ContentUnderstandingClient | None = None,
+        analyzer_id: str | None = None,
+        max_wait: float | None = DEFAULT_MAX_WAIT_SECONDS,
+        output_sections: list[AnalysisSection] | None = None,
+        file_search: FileSearchConfig | None = None,
+        source_id: str = DEFAULT_SOURCE_ID,
+        env_file_path: str | None = None,
+        env_file_encoding: str | None = None,
+    ) -> None:
+        super().__init__(source_id)
+
+        if client is not None:
+            # Use the pre-built client directly — endpoint/credential are ignored.
+            self._client = client
+            self._owns_client = False
+            self._endpoint = ""
+            self._credential = None
+        else:
+            # Build a new client from endpoint + credential.
+            settings = load_settings(
+                ContentUnderstandingSettings,
+                env_prefix="AZURE_CONTENTUNDERSTANDING_",
+                required_fields=["endpoint"],
+                endpoint=endpoint,
+                env_file_path=env_file_path,
+                env_file_encoding=env_file_encoding,
+            )
+            resolved_endpoint: str = settings["endpoint"]  # type: ignore[assignment]  # validated by load_settings
+
+            if credential is None:
+                raise ValueError(
+                    "Azure credential is required. Provide a 'credential' keyword argument "
+                    "(e.g., AzureKeyCredential or AzureCliCredential), or pass a pre-built "
+                    "'client' (ContentUnderstandingClient) instead."
+                )
+
+            self._endpoint = resolved_endpoint
+            self._credential = credential
+            self._client = ContentUnderstandingClient(
+                self._endpoint, self._credential, user_agent=AGENT_FRAMEWORK_USER_AGENT
+            )
+            self._owns_client = True
+        self.analyzer_id = analyzer_id
+        self.max_wait = max_wait
+        self.output_sections: list[AnalysisSection] = output_sections or ["markdown", "fields"]
+        self.file_search = file_search
+        # Global list of uploaded file IDs — used only by close() for
+        # best-effort cleanup.  The authoritative per-session copy lives in
+        # state["_uploaded_file_ids"] (populated in before_run).  This global
+        # list may contain entries from multiple sessions; that is intentional
+        # for cleanup.
+        self._all_uploaded_file_ids: list[str] = []
+
+    async def __aenter__(self) -> Self:
+        """Async context manager entry."""
+        return self
+
+    async def __aexit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: Any,
+    ) -> None:
+        """Async context manager exit — cleanup clients."""
+        await self.close()
+
+    async def close(self) -> None:
+        """Close the underlying CU client and clean up resources.
+
+        Uses global tracking lists for best-effort cleanup across all
+        sessions that used this provider instance.
+        """
+        # Clean up uploaded files; the vector store itself is caller-managed.
+        if self.file_search and self._all_uploaded_file_ids:
+            await self._cleanup_uploaded_files()
+        # Only close the client if we created it internally.
+        # When a pre-built client was passed in, the caller owns its lifecycle.
+        if self._owns_client:
+            await self._client.close()
+
+    async def before_run(
+        self,
+        *,
+        agent: SupportsAgentRun,
+        session: AgentSession,
+        context: SessionContext,
+        state: dict[str, Any],
+    ) -> None:
+        """Analyze file attachments and inject results into the LLM context.
+
+        This method is called automatically by the framework before each LLM invocation.
+        """
+        documents: dict[str, DocumentEntry] = state.setdefault("documents", {})
+
+        # Per-session mutable state — isolated per session to prevent cross-session leakage.
+        # _pending_tokens stores serializable continuation tokens (not asyncio.Task objects)
+        # so that state can be persisted to disk/storage by the framework.
+        # Structure: {doc_key: {"continuation_token": <opaque Azure SDK string>,
+        #                       "analyzer_id": <CU analyzer used for this file>}}
+        pending_tokens: dict[str, dict[str, str]] = state.setdefault("_pending_tokens", {})
+        pending_uploads: list[tuple[str, DocumentEntry]] = state.setdefault("_pending_uploads", [])
+
+        # Resolve pending Content Understanding analysis from its continuation tokens
+        await self._resolve_pending_analysis(pending_tokens, pending_uploads, documents, context)
+
+        # 1b. Upload any documents that completed in the background (file_search mode)
+        if pending_uploads:
+            # Use a bounded timeout so before_run() stays responsive and does not block
+            # indefinitely on slow vector store indexing.
+            upload_timeout = getattr(self, "max_wait", None)
+            remaining_uploads: list[tuple[str, DocumentEntry]] = []
+            for upload_key, upload_entry in pending_uploads:
+                try:
+                    if upload_timeout is not None:
+                        await asyncio.wait_for(
+                            self._upload_to_vector_store(upload_key, upload_entry, state=state),
+                            timeout=upload_timeout,
+                        )
+                    else:
+                        await self._upload_to_vector_store(upload_key, upload_entry, state=state)
+                except asyncio.TimeoutError:
+                    # Leave timed-out uploads pending so they can be retried on a later turn.
+                    logger.warning(
+                        "Timed out while uploading document '%s' to vector store; will retry later.",
+                        upload_key,
+                    )
+                    remaining_uploads.append((upload_key, upload_entry))
+                except Exception:
+                    # Log unexpected failures and drop the upload entry; this matches prior
+                    # behavior where all pending uploads were cleared regardless of outcome.
+                    logger.exception(
+                        "Error while uploading document '%s' to vector store; dropping from pending list.",
+                        upload_key,
+                    )
+                    context.extend_messages(
+                        self.source_id,
+                        [
+                            Message(
+                                role="user",
+                                contents=[
+                                    (
+                                        f"Document '{upload_key}' was analyzed but failed to upload "
+                                        "to the vector store. The document content is not available for search."
+                                    )
+                                ],
+                            )
+                        ],
+                    )
+            state["_pending_uploads"] = remaining_uploads
+            pending_uploads = remaining_uploads
+
+        # 2. Detect CU-supported file attachments, strip them from input, and return for analysis
+        new_files = detect_and_strip_files(context)
+
+        # 3. Analyze new files using CU (track elapsed time for combined timeout)
+        file_start_times: dict[str, float] = {}
+        accepted_keys: set[str] = set()  # doc_keys successfully accepted for analysis this turn
+        for doc_key, content_item, binary_data in new_files:
+            # Reject duplicate filenames — re-analyzing would orphan vector store entries
+            if doc_key in documents:
+                logger.warning("Duplicate document key '%s' — skipping (already exists in session).", doc_key)
+                context.extend_messages(
+                    self.source_id,
+                    [
+                        Message(
+                            role="user",
+                            contents=[
+                                (
+                                    f"The user tried to upload '{doc_key}', but a file with that name "
+                                    "was already uploaded earlier in this session. The new upload was rejected "
+                                    "and was not analyzed. Tell the user that a file with the same name "
+                                    "already exists and they need to rename the file before uploading again."
+                                )
+                            ],
+                        )
+                    ],
+                )
+                continue
+            file_start_times[doc_key] = time.monotonic()
+            doc_entry = await self._analyze_file(doc_key, content_item, binary_data, context, pending_tokens)
+            if doc_entry:
+                documents[doc_key] = doc_entry
+                accepted_keys.add(doc_key)
+
+        # 4. Inject content for ready documents and register tools
+        if documents:
+            self._register_tools(documents, context)
+
+        # 5. On upload turns, inject content for docs accepted this turn
+        for doc_key in accepted_keys:
+            entry = documents.get(doc_key)
+            if entry and entry["status"] == DocumentStatus.READY and entry["result"]:
+                # Upload to vector store if file_search is configured
+                if self.file_search:
+                    # Combined timeout: subtract CU analysis time from max_wait
+                    remaining: float | None = None
+                    if self.max_wait is not None:
+                        elapsed = time.monotonic() - file_start_times.get(doc_key, time.monotonic())
+                        remaining = max(0.0, self.max_wait - elapsed)
+                    uploaded = await self._upload_to_vector_store(doc_key, entry, timeout=remaining, state=state)
+                    if uploaded:
+                        context.extend_messages(
+                            self.source_id,
+                            [
+                                Message(
+                                    role="user",
+                                    contents=[
+                                        (
+                                            f"The user just uploaded '{entry['filename']}'. It has been analyzed "
+                                            "using Azure Content Understanding and indexed in a vector store. "
+                                            f"When using file_search, include '{entry['filename']}' in your query "
+                                            "to retrieve content from this specific document."
+                                        )
+                                    ],
+                                )
+                            ],
+                        )
+                    elif entry.get("error"):
+                        # Upload failed (not timeout — actual error)
+                        context.extend_messages(
+                            self.source_id,
+                            [
+                                Message(
+                                    role="user",
+                                    contents=[
+                                        (
+                                            f"Document '{entry['filename']}' was analyzed but failed to upload "
+                                            "to the vector store. The document content is not available for search."
+                                        )
+                                    ],
+                                )
+                            ],
+                        )
+                    else:
+                        # Upload deferred to background (timeout)
+                        context.extend_messages(
+                            self.source_id,
+                            [
+                                Message(
+                                    role="user",
+                                    contents=[
+                                        (
+                                            f"Document '{entry['filename']}' has been analyzed and is being indexed. "
+                                            "Ask about it again in a moment."
+                                        )
+                                    ],
+                                )
+                            ],
+                        )
+                else:
+                    # Without file_search, inject full content into context
+                    context.extend_messages(
+                        self,
+                        [
+                            Message(role="user", contents=[entry["result"] or ""]),
+                        ],
+                    )
+                    context.extend_messages(
+                        self.source_id,
+                        [
+                            Message(
+                                role="user",
+                                contents=[
+                                    (
+                                        f"The user just uploaded '{entry['filename']}'."
+                                        " It has been analyzed using Azure Content Understanding."
+                                        " The document content (markdown) and extracted fields"
+                                        " (YAML front matter) are provided above."
+                                        " If the user's question is ambiguous,"
+                                        " prioritize this most recently uploaded document."
+                                        " Use specific field values and cite page numbers"
+                                        " when answering."
+                                    )
+                                ],
+                            )
+                        ],
+                    )
+
+        # 6. Register file_search tool (for LLM clients that support it)
+        if self.file_search:
+            context.extend_tools(
+                self.source_id,
+                [self.file_search.file_search_tool],
+            )
+            context.extend_instructions(
+                self.source_id,
+                "Tool usage guidelines:\n"
+                "- Use file_search ONLY when answering questions about document content.\n"
+                "- Use list_documents() for status queries (e.g. 'list docs', 'what's uploaded?').\n"
+                "- Do NOT call file_search for status queries — it wastes tokens.",
+            )
+
+    # ------------------------------------------------------------------
+    # Analyzer Resolution
+    # ------------------------------------------------------------------
+
+    def _resolve_analyzer_id(self, media_type: str) -> str:
+        """Return the analyzer ID to use for the given media type.
+
+        When ``self.analyzer_id`` is set, it is always returned (explicit
+        override).  Otherwise the media type prefix is matched against the
+        known mapping, falling back to ``prebuilt-documentSearch``.
+        """
+        if self.analyzer_id is not None:
+            return self.analyzer_id
+        for prefix, analyzer in MEDIA_TYPE_ANALYZER_MAP.items():
+            if media_type.startswith(prefix):
+                return analyzer
+        return DEFAULT_ANALYZER
+
+    # ------------------------------------------------------------------
+    # Analysis
+    # ------------------------------------------------------------------
+
+    async def _analyze_file(
+        self,
+        doc_key: str,
+        content: Content,
+        binary_data: bytes | None,
+        context: SessionContext,
+        pending_tokens: dict[str, dict[str, str]] | None = None,
+    ) -> DocumentEntry | None:
+        """Analyze a single file via CU with timeout handling.
+
+        The analyzer is resolved in priority order:
+        1. Per-file override via ``content.additional_properties["analyzer_id"]``
+        2. Provider-level default via ``self.analyzer_id``
+        3. Auto-detect by media type (document/audio/video)
+
+        Returns:
+            A ``DocumentEntry`` (ready, analyzing, or failed), or ``None`` if
+            file data could not be extracted.
+        """
+        media_type = content.media_type or "application/octet-stream"
+        filename = doc_key
+
+        # Per-file analyzer override from additional_properties
+        props = content.additional_properties or {}
+        per_file_analyzer = props.get("analyzer_id")
+        content_range = props.get("content_range")
+        resolved_analyzer = per_file_analyzer or self._resolve_analyzer_id(media_type)
+        t0 = time.monotonic()
+
+        try:
+            # Start CU analysis
+            if content.type == "uri" and content.uri and not content.uri.startswith("data:"):
+                poller = await self._client.begin_analyze(
+                    resolved_analyzer,
+                    inputs=[AnalysisInput(url=content.uri, content_range=content_range)],
+                )
+            elif binary_data:
+                poller = await self._client.begin_analyze_binary(
+                    resolved_analyzer,
+                    binary_input=binary_data,
+                    content_type=media_type,
+                )
+            else:
+                context.extend_messages(
+                    self.source_id,
+                    [Message(role="user", contents=[f"Could not extract file data from '{filename}'."])],
+                )
+                return None
+
+            # Wait with timeout; defer to background polling on timeout.
+            try:
+                result = await asyncio.wait_for(poller.result(), timeout=self.max_wait)
+            except asyncio.TimeoutError:
+                # Save continuation token for resuming on next before_run().
+                # Continuation tokens are serializable strings, so state can
+                # be persisted to disk/storage without issues.
+                token = poller.continuation_token()
+                logger.info("Analysis of '%s' timed out; deferring to background via continuation token.", filename)
+                if pending_tokens is not None:
+                    pending_tokens[doc_key] = {
+                        "continuation_token": token,
+                        "analyzer_id": resolved_analyzer,
+                    }
+                context.extend_messages(
+                    self.source_id,
+                    [
+                        Message(
+                            role="user",
+                            contents=[f"Document '{filename}' is being analyzed. Ask about it again in a moment."],
+                        )
+                    ],
+                )
+                return DocumentEntry(
+                    status=DocumentStatus.ANALYZING,
+                    filename=filename,
+                    media_type=media_type,
+                    analyzer_id=resolved_analyzer,
+                    analyzed_at=None,
+                    analysis_duration_s=None,
+                    upload_duration_s=None,
+                    result=None,
+                    error=None,
+                )
+
+            # Analysis completed within timeout
+            analysis_duration = round(time.monotonic() - t0, 2)
+            rendered = self._render_for_llm(result, filename)
+            logger.info("Analyzed '%s' with analyzer '%s' in %.1fs.", filename, resolved_analyzer, analysis_duration)
+            return DocumentEntry(
+                status=DocumentStatus.READY,
+                filename=filename,
+                media_type=media_type,
+                analyzer_id=resolved_analyzer,
+                analyzed_at=datetime.now(tz=timezone.utc).isoformat(),
+                analysis_duration_s=analysis_duration,
+                upload_duration_s=None,
+                result=rendered,
+                error=None,
+            )
+
+        except asyncio.TimeoutError:
+            raise
+        except Exception as e:
+            logger.warning("CU analysis error for '%s': %s", filename, e)
+            context.extend_messages(
+                self.source_id,
+                [Message(role="user", contents=[f"Could not analyze '{filename}': {e}"])],
+            )
+            return DocumentEntry(
+                status=DocumentStatus.FAILED,
+                filename=filename,
+                media_type=media_type,
+                analyzer_id=resolved_analyzer,
+                analyzed_at=datetime.now(tz=timezone.utc).isoformat(),
+                analysis_duration_s=round(time.monotonic() - t0, 2),
+                upload_duration_s=None,
+                result=None,
+                error=str(e),
+            )
+
+    # ------------------------------------------------------------------
+    # Pending Analysis Resolution
+    # ------------------------------------------------------------------
+
+    async def _resolve_pending_analysis(
+        self,
+        pending_tokens: dict[str, dict[str, str]],
+        pending_uploads: list[tuple[str, DocumentEntry]],
+        documents: dict[str, DocumentEntry],
+        context: SessionContext,
+    ) -> None:
+        """Resume pending CU analyses using serializable continuation tokens.
+
+        When a file's CU analysis exceeds ``max_wait``, a continuation token
+        (an opaque string from the Azure SDK) is saved in ``state`` instead of
+        an ``asyncio.Task``.  This keeps state fully serializable — it can be
+        persisted to disk/storage by the framework.
+
+        On the next ``before_run()`` call, this method resumes each pending
+        operation by passing the token back to ``begin_analyze()``.  If the
+        server-side operation has completed, the result is available
+        immediately; otherwise the token is kept for the next turn.
+        """
+        if not pending_tokens:
+            return
+        logger.info("Resolving %d pending analysis token(s).", len(pending_tokens))
+        completed_keys: list[str] = []
+
+        for doc_key, token_info in pending_tokens.items():
+            entry = documents.get(doc_key)
+            if not entry:
+                completed_keys.append(doc_key)
+                continue
+
+            try:
+                poller = await self._client.begin_analyze(  # type: ignore[call-overload, reportUnknownVariableType]
+                    token_info["analyzer_id"],
+                    continuation_token=token_info["continuation_token"],
+                )
+                # Use wait_for to avoid blocking before_run indefinitely.
+                # poller.done() always returns False for resumed pollers (stale
+                # cached status), so we call poller.result() which polls the server.
+                #
+                # Timeout: at least 10s regardless of max_wait.  The upload-turn
+                # max_wait can be very short (e.g. 5s) for responsiveness, but
+                # on resolution turns the resumed poller needs a network round-trip
+                # to fetch the result.  If the analysis is still running after 10s,
+                # the token is kept and retried on the next turn.
+                MIN_RESOLUTION_TIMEOUT = 10.0
+                resolution_timeout = max(self.max_wait or MIN_RESOLUTION_TIMEOUT, MIN_RESOLUTION_TIMEOUT)
+                try:
+                    result: AnalysisResult = await asyncio.wait_for(
+                        poller.result(),  # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType]
+                        timeout=resolution_timeout,
+                    )
+                except asyncio.TimeoutError:
+                    # Still running — update token and keep for next turn
+                    new_token: str = poller.continuation_token()  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]
+                    token_info["continuation_token"] = new_token
+                    logger.info("Analysis for '%s' still running; keeping token for next turn.", doc_key)
+                    continue
+
+                completed_keys.append(doc_key)
+                rendered = self._render_for_llm(result, entry["filename"])
+                entry["status"] = DocumentStatus.READY
+                entry["analyzed_at"] = datetime.now(tz=timezone.utc).isoformat()
+                entry["result"] = rendered
+                entry["error"] = None
+                logger.info("Background analysis of '%s' completed.", entry["filename"])
+
+                # Inject newly ready content
+                if self.file_search:
+                    pending_uploads.append((doc_key, entry))
+                else:
+                    context.extend_messages(
+                        self,
+                        [
+                            Message(role="user", contents=[rendered]),
+                        ],
+                    )
+                context.extend_messages(
+                    self.source_id,
+                    [
+                        Message(
+                            role="user",
+                            contents=[
+                                f"Document '{entry['filename']}' analysis is now complete."
+                                + (
+                                    " The document is being indexed in the vector store and will become"
+                                    " searchable via file_search shortly."
+                                    if self.file_search
+                                    else " The content is provided above."
+                                )
+                            ],
+                        )
+                    ],
+                )
+
+            except Exception as e:
+                completed_keys.append(doc_key)
+                logger.warning("Background analysis of '%s' failed: %s", entry.get("filename", doc_key), e)
+                entry["status"] = DocumentStatus.FAILED
+                entry["analyzed_at"] = datetime.now(tz=timezone.utc).isoformat()
+                entry["error"] = str(e)
+                context.extend_messages(
+                    self.source_id,
+                    [Message(role="user", contents=[f"Document '{entry['filename']}' analysis failed: {e}"])],
+                )
+
+        for key in completed_keys:
+            del pending_tokens[key]
+
+    # ------------------------------------------------------------------
+    # LLM Input Rendering (delegates to azure.ai.contentunderstanding.to_llm_input)
+    # ------------------------------------------------------------------
+
+    def _render_for_llm(
+        self,
+        result: AnalysisResult,
+        filename: str,
+    ) -> str:
+        """Render a CU ``AnalysisResult`` into LLM-friendly text.
+
+        Maps the MAF ``output_sections`` list to ``to_llm_input`` kwargs:
+
+        - ``"markdown" in output_sections`` -> ``include_markdown=True``
+        - ``"fields" in output_sections``  -> ``include_fields=True``
+
+        Args:
+            result: The CU analysis result.
+            filename: Document filename, surfaced to the LLM via the
+                ``source`` front matter key.
+
+        Returns:
+            A YAML-front-matter-prefixed text block ready for direct LLM
+            consumption or vector store upload.
+        """
+        return to_llm_input(
+            result,
+            include_markdown="markdown" in self.output_sections,
+            include_fields="fields" in self.output_sections,
+            metadata={"source": filename},
+        )
+
+    # ------------------------------------------------------------------
+    # Tool Registration
+    # ------------------------------------------------------------------
+
+    def _register_tools(
+        self,
+        documents: dict[str, DocumentEntry],
+        context: SessionContext,
+    ) -> None:
+        """Register document tools on the context.
+
+        Only ``list_documents`` is registered — the full document content is
+        already injected into conversation history on the upload turn, so a
+        separate retrieval tool is not needed.
+        """
+        context.extend_tools(
+            self.source_id,
+            [self._make_list_documents_tool(documents)],
+        )
+
+    @staticmethod
+    def _make_list_documents_tool(documents: dict[str, DocumentEntry]) -> FunctionTool:
+        """Create a tool that lists all tracked documents with their status."""
+        docs_ref = documents
+
+        def list_documents() -> str:
+            """List all documents that have been uploaded and their analysis status."""
+            entries: list[dict[str, object]] = []
+            for name, entry in docs_ref.items():
+                entries.append({
+                    "name": name,
+                    "status": entry["status"],
+                    "media_type": entry["media_type"],
+                    "analyzed_at": entry["analyzed_at"],
+                    "analysis_duration_s": entry["analysis_duration_s"],
+                    "upload_duration_s": entry["upload_duration_s"],
+                })
+            return json.dumps(entries, indent=2, default=str)
+
+        return FunctionTool(
+            name="list_documents",
+            description=(
+                "List all documents that have been uploaded in this session "
+                "with their analysis status (analyzing, uploading, ready, or failed)."
+            ),
+            func=list_documents,
+        )
+
+    # ------------------------------------------------------------------
+    # file_search Vector Store Integration
+    # ------------------------------------------------------------------
+
+    async def _upload_to_vector_store(
+        self,
+        doc_key: str,
+        entry: DocumentEntry,
+        *,
+        timeout: float | None = None,
+        state: dict[str, Any] | None = None,
+    ) -> bool:
+        """Upload CU-extracted markdown to the caller's vector store.
+
+        Delegates to the configured ``FileSearchBackend`` (OpenAI, Foundry,
+        or a custom implementation). The upload includes file upload **and**
+        vector store indexing (embedding + ingestion) — ``create_and_poll``
+        waits for the index to be fully ready before returning.
+
+        Args:
+            doc_key: Document identifier.
+            entry: The document entry with extracted results.
+            timeout: Max seconds to wait for upload + indexing. ``None`` waits
+                indefinitely. On timeout the upload is deferred to the
+                per-session ``_pending_uploads`` queue for the next
+                ``before_run()`` call.
+            state: Per-session state dict for tracking uploaded file IDs and
+                pending uploads.
+
+        Returns:
+            True if the upload succeeded, False otherwise.
+        """
+        if not self.file_search:
+            return False
+
+        result = entry.get("result")
+        if not result:
+            return False
+
+        if not _has_renderable_body(result):
+            # Empty CU result (e.g. blank markdown, no fields) — skip the
+            # upload so the vector store stays clean. The DocumentEntry still
+            # records the front-matter-only ``result`` so callers can introspect.
+            return False
+
+        entry["status"] = DocumentStatus.UPLOADING
+        t0 = time.monotonic()
+
+        try:
+            upload_coro = self.file_search.backend.upload_file(
+                self.file_search.vector_store_id, f"{doc_key}.md", result.encode("utf-8")
+            )
+            file_id = await asyncio.wait_for(upload_coro, timeout=timeout)
+            upload_duration = round(time.monotonic() - t0, 2)
+            # Track in per-session state and global list (for close() cleanup)
+            if state is not None:
+                state.setdefault("_uploaded_file_ids", []).append(file_id)
+            self._all_uploaded_file_ids.append(file_id)
+            entry["status"] = DocumentStatus.READY
+            entry["upload_duration_s"] = upload_duration
+            logger.info("Uploaded '%s' to vector store in %.1fs (%s bytes).", doc_key, upload_duration, len(result))
+            return True
+
+        except asyncio.TimeoutError:
+            logger.info("Vector store upload for '%s' timed out; deferring to background.", doc_key)
+            entry["status"] = DocumentStatus.UPLOADING
+            if state is not None:
+                state.setdefault("_pending_uploads", []).append((doc_key, entry))
+            return False
+
+        except Exception as e:
+            logger.warning("Failed to upload '%s' to vector store: %s", doc_key, e)
+            entry["status"] = DocumentStatus.FAILED
+            entry["upload_duration_s"] = round(time.monotonic() - t0, 2)
+            entry["error"] = f"Vector store upload failed: {e}"
+            return False
+
+    async def _cleanup_uploaded_files(self) -> None:
+        """Delete files uploaded by this provider via the configured backend.
+
+        The vector store itself is caller-managed and is not deleted here.
+        """
+        if not self.file_search:
+            return
+
+        backend = self.file_search.backend
+
+        try:
+            for file_id in self._all_uploaded_file_ids:
+                await backend.delete_file(file_id)
+            self._all_uploaded_file_ids.clear()
+
+        except Exception as e:
+            logger.warning("Failed to clean up uploaded files: %s", e)

python/packages/azure-contentunderstanding/agent_framework_azure_contentunderstanding/_detection.py

@@ -0,0 +1,234 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""File detection utilities for Azure Content Understanding context provider.
+
+Functions for scanning input messages, sniffing MIME types, deriving
+document keys, and extracting binary data from content items.
+"""
+
+from __future__ import annotations
+
+import base64
+import logging
+import mimetypes
+import re
+import uuid
+
+import filetype
+from agent_framework import Content, SessionContext
+
+logger = logging.getLogger("agent_framework.azure_contentunderstanding")
+
+# MIME types used to match against the resolved media type for routing files to CU analysis.
+# The media type may be provided via Content.media_type or inferred (e.g., via sniffing or filename)
+# when missing or generic (such as application/octet-stream). Only files whose resolved media type is
+# in this set will be processed; others are skipped.
+#
+# Supported input file types:
+# https://learn.microsoft.com/azure/ai-services/content-understanding/service-limits#input-file-limits
+SUPPORTED_MEDIA_TYPES: frozenset[str] = frozenset({
+    # Documents and images
+    "application/pdf",
+    "image/jpeg",
+    "image/png",
+    "image/tiff",
+    "image/bmp",
+    "image/heif",
+    "image/heic",
+    "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
+    "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
+    "application/vnd.openxmlformats-officedocument.presentationml.presentation",
+    # Text
+    "text/plain",
+    "text/html",
+    "text/markdown",
+    "text/rtf",
+    "text/xml",
+    "application/xml",
+    "message/rfc822",
+    "application/vnd.ms-outlook",
+    # Audio
+    "audio/wav",
+    "audio/mpeg",
+    "audio/mp3",
+    "audio/mp4",
+    "audio/m4a",
+    "audio/flac",
+    "audio/ogg",
+    "audio/opus",
+    "audio/webm",
+    "audio/x-ms-wma",
+    "audio/aac",
+    "audio/amr",
+    "audio/3gpp",
+    # Video
+    "video/mp4",
+    "video/quicktime",
+    "video/x-msvideo",
+    "video/webm",
+    "video/x-flv",
+    "video/x-ms-wmv",
+    "video/x-ms-asf",
+    "video/x-matroska",
+})
+
+# Mapping from filetype's MIME output to our canonical SUPPORTED_MEDIA_TYPES values.
+# filetype uses some x-prefixed variants that differ from our set.
+MIME_ALIASES: dict[str, str] = {
+    "audio/x-wav": "audio/wav",
+    "audio/x-flac": "audio/flac",
+    "video/x-m4v": "video/mp4",
+}
+
+
+def detect_and_strip_files(
+    context: SessionContext,
+) -> list[tuple[str, Content, bytes | None]]:
+    """Scan input messages for supported file content and prepare for CU analysis.
+
+    Scans for type ``data`` or ``uri`` content supported by Azure Content
+    Understanding, strips them from messages to prevent raw binary being sent
+    to the LLM, and returns metadata for CU analysis.
+
+    Detected files are tracked via ``doc_key`` (derived from filename, URL,
+    or UUID) and their analysis status is managed in session state.
+
+    When the upstream MIME type is unreliable (``application/octet-stream``
+    or missing), binary content sniffing via ``filetype`` is used to
+    determine the real media type, with ``mimetypes.guess_type`` as a
+    filename-based fallback.
+
+    Returns:
+        List of (doc_key, content_item, binary_data) tuples for files to analyze.
+    """
+    results: list[tuple[str, Content, bytes | None]] = []
+    strip_ids: set[int] = set()
+
+    for msg in context.input_messages:
+        for c in msg.contents:
+            if c.type not in ("data", "uri"):
+                continue
+
+            media_type = c.media_type
+            # Fast path: already a known supported type
+            if media_type and media_type in SUPPORTED_MEDIA_TYPES:
+                binary_data = extract_binary(c)
+                results.append((derive_doc_key(c), c, binary_data))
+                strip_ids.add(id(c))
+                continue
+
+            # Slow path: unreliable MIME — sniff binary content
+            if (not media_type) or (media_type == "application/octet-stream"):
+                binary_data = extract_binary(c)
+                resolved = sniff_media_type(binary_data, c)
+                if resolved and (resolved in SUPPORTED_MEDIA_TYPES):
+                    c.media_type = resolved
+                    results.append((derive_doc_key(c), c, binary_data))
+                    strip_ids.add(id(c))
+
+        # Strip detected files from input so raw binary isn't sent to LLM
+        msg.contents = [c for c in msg.contents if id(c) not in strip_ids]
+
+    return results
+
+
+def sniff_media_type(binary_data: bytes | None, content: Content) -> str | None:
+    """Sniff the actual MIME type from binary data, with filename fallback.
+
+    Uses ``filetype`` (magic-bytes) first, then ``mimetypes.guess_type``
+    on the filename. Normalizes filetype's variant MIME values (e.g.
+    ``audio/x-wav`` -> ``audio/wav``) via ``MIME_ALIASES``.
+    """
+    # 1. Binary sniffing via filetype (needs only first 261 bytes)
+    if binary_data:
+        kind = filetype.guess(binary_data[:262])  # type: ignore[reportUnknownMemberType]
+        if kind:
+            mime: str = kind.mime
+            return MIME_ALIASES.get(mime, mime)
+
+    # 2. Filename extension fallback — try additional_properties first,
+    # then extract basename from external URL path
+    filename: str | None = None
+    if content.additional_properties:
+        filename = content.additional_properties.get("filename")
+    if not filename and content.uri and not content.uri.startswith("data:"):
+        # Extract basename from URL path (e.g. "https://example.com/report.pdf?v=1" -> "report.pdf")
+        filename = content.uri.split("?")[0].split("#")[0].rsplit("/", 1)[-1]
+    if filename:
+        guessed, _ = mimetypes.guess_type(filename)  # uses file extension to guess MIME type
+        if guessed:
+            return MIME_ALIASES.get(guessed, guessed)
+
+    return None
+
+
+def is_supported_content(content: Content) -> bool:
+    """Check if a content item is a supported file type for CU analysis."""
+    if content.type not in ("data", "uri"):
+        return False
+    media_type = content.media_type
+    if not media_type:
+        return False
+    return media_type in SUPPORTED_MEDIA_TYPES
+
+
+def sanitize_doc_key(raw: str) -> str:
+    """Sanitize a document key to prevent prompt injection.
+
+    Removes control characters (newlines, tabs, etc.), collapses
+    whitespace, strips surrounding whitespace, and caps length at
+    255 characters.
+    """
+    # Remove control characters (C0/C1 controls, including \n, \r, \t)
+    cleaned = re.sub(r"[\x00-\x1f\x7f-\x9f]", "", raw)
+    # Collapse whitespace
+    cleaned = " ".join(cleaned.split())
+    # Cap length
+    return cleaned[:255] if cleaned else f"doc_{uuid.uuid4().hex[:8]}"
+
+
+def derive_doc_key(content: Content) -> str:
+    """Derive a unique document key from content metadata.
+
+    The key is used to track documents in session state. Duplicate keys
+    within a session are rejected (not re-analyzed) to prevent orphaned
+    vector store entries.
+
+    The returned key is sanitized to prevent prompt injection via
+    crafted filenames (control characters removed, length capped).
+
+    Priority: filename > URL basename > generated UUID.
+    """
+    # 1. Filename from additional_properties
+    if content.additional_properties:
+        filename = content.additional_properties.get("filename")
+        if filename and isinstance(filename, str):
+            return sanitize_doc_key(filename)
+
+    # 2. URL path basename for external URIs (e.g. "https://example.com/report.pdf" -> "report.pdf")
+    if content.type == "uri" and content.uri and not content.uri.startswith("data:"):
+        path = content.uri.split("?")[0].split("#")[0]  # strip query params and fragments
+        # rstrip("/") handles trailing slashes (e.g. ".../files/" -> ".../files")
+        # rsplit("/", 1)[-1] splits from the right once to get the last path segment
+        basename = path.rstrip("/").rsplit("/", 1)[-1]
+        if basename:
+            return sanitize_doc_key(basename)
+
+    # 3. Fallback: generate a unique ID for anonymous uploads (no filename, no URL)
+    return f"doc_{uuid.uuid4().hex[:8]}"
+
+
+def extract_binary(content: Content) -> bytes | None:
+    """Extract binary data from a data URI content item.
+
+    Only handles ``data:`` URIs (base64-encoded). Returns ``None`` for
+    external URLs -- those are passed directly to CU via ``begin_analyze``.
+    """
+    if content.uri and content.uri.startswith("data:"):
+        try:
+            _, data_part = content.uri.split(",", 1)
+            return base64.b64decode(data_part)
+        except Exception:
+            logger.warning("Failed to decode base64 data URI")
+            return None
+    return None

python/packages/azure-contentunderstanding/agent_framework_azure_contentunderstanding/_file_search.py

@@ -0,0 +1,101 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""File search backend abstraction for vector store file operations.
+
+Provides a unified interface for uploading CU-extracted content to
+vector stores across different LLM clients. Two implementations:
+
+- ``OpenAIFileSearchBackend`` — for ``OpenAIChatClient`` (Responses API)
+- ``FoundryFileSearchBackend`` — for ``FoundryChatClient`` (Responses API via Azure)
+
+Both share the same OpenAI-compatible vector store file API but differ
+in the file upload ``purpose`` value.
+
+Vector store creation, tool construction, and lifecycle management are
+the caller's responsibility — the backend only handles file upload/delete.
+"""
+
+from __future__ import annotations
+
+import io
+from abc import ABC, abstractmethod
+from typing import Any
+
+
+class FileSearchBackend(ABC):
+    """Abstract interface for vector store file operations.
+
+    Implementations handle the differences between OpenAI and Foundry
+    file upload APIs (e.g., different ``purpose`` values).
+
+    Vector store creation, deletion, and ``file_search`` tool construction
+    are **not** part of this interface — those are managed by the caller.
+    """
+
+    @abstractmethod
+    async def upload_file(self, vector_store_id: str, filename: str, content: bytes) -> str:
+        """Upload a file to a vector store and return the file ID."""
+
+    @abstractmethod
+    async def delete_file(self, file_id: str) -> None:
+        """Delete a previously uploaded file by ID."""
+
+
+class _OpenAICompatBackend(FileSearchBackend):
+    """Shared base for OpenAI-compatible file upload backends.
+
+    Both OpenAI and Foundry use the same ``client.files.*`` and
+    ``client.vector_stores.files.*`` API surface. Subclasses only
+    override the file upload ``purpose``.
+    """
+
+    _FILE_PURPOSE: str  # Subclasses must set this
+
+    def __init__(self, client: Any) -> None:
+        self._client = client
+
+    async def upload_file(self, vector_store_id: str, filename: str, content: bytes) -> str:
+        uploaded = await self._client.files.create(
+            file=(filename, io.BytesIO(content)),
+            purpose=self._FILE_PURPOSE,
+        )
+        # Use create_and_poll to wait for indexing to complete before returning.
+        # Without this, file_search queries may return no results immediately
+        # after upload because the vector store index isn't ready yet.
+        await self._client.vector_stores.files.create_and_poll(
+            vector_store_id=vector_store_id,
+            file_id=uploaded.id,
+        )
+        return uploaded.id
+
+    async def delete_file(self, file_id: str) -> None:
+        await self._client.files.delete(file_id)
+
+
+class OpenAIFileSearchBackend(_OpenAICompatBackend):
+    """File search backend for OpenAI Responses API.
+
+    Use with ``OpenAIChatClient`` or ``AzureOpenAIResponsesClient``.
+    Requires an ``AsyncOpenAI`` or ``AsyncAzureOpenAI`` client.
+
+    Args:
+        client: An async OpenAI client (``AsyncOpenAI`` or ``AsyncAzureOpenAI``)
+            that supports ``client.files.*`` and ``client.vector_stores.*`` APIs.
+    """
+
+    _FILE_PURPOSE = "user_data"
+
+
+class FoundryFileSearchBackend(_OpenAICompatBackend):
+    """File search backend for Azure AI Foundry.
+
+    Use with ``FoundryChatClient``. Requires the OpenAI-compatible client
+    obtained from ``FoundryChatClient.client`` (i.e.,
+    ``project_client.get_openai_client()``).
+
+    Args:
+        client: The OpenAI-compatible async client from a ``FoundryChatClient``
+            (access via ``foundry_client.client``).
+    """
+
+    _FILE_PURPOSE = "assistants"

python/packages/azure-contentunderstanding/agent_framework_azure_contentunderstanding/_models.py

@@ -0,0 +1,121 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import Enum
+from typing import Any, Literal, TypedDict
+
+from ._file_search import FileSearchBackend, FoundryFileSearchBackend, OpenAIFileSearchBackend
+
+
+class DocumentStatus(str, Enum):
+    """Analysis lifecycle state of a tracked document."""
+
+    ANALYZING = "analyzing"
+    """CU analysis is in progress (deferred to background)."""
+
+    UPLOADING = "uploading"
+    """Analysis complete; vector store upload + indexing is in progress."""
+
+    READY = "ready"
+    """Analysis (and upload, if applicable) completed successfully."""
+
+    FAILED = "failed"
+    """Analysis or upload failed."""
+
+
+AnalysisSection = Literal["markdown", "fields"]
+"""Which sections of the CU output to pass to the LLM.
+
+- ``"markdown"``: Full document text with tables as HTML, reading order preserved.
+- ``"fields"``: Extracted typed fields with confidence scores (when available).
+"""
+
+
+class DocumentEntry(TypedDict):
+    """Tracks the analysis state of a single document in session state."""
+
+    status: DocumentStatus
+    filename: str
+    media_type: str
+    analyzer_id: str
+    analyzed_at: str | None
+    analysis_duration_s: float | None
+    upload_duration_s: float | None
+    result: str | None
+    """LLM-ready text rendered by ``azure.ai.contentunderstanding.to_llm_input``.
+
+    Stored as a string (YAML front matter + markdown body) so every consumer
+    (LLM context injection, vector store upload) can use it without re-rendering.
+    ``None`` until analysis completes successfully.
+    """
+    error: str | None
+
+
+@dataclass
+class FileSearchConfig:
+    """Configuration for uploading CU-extracted content to an existing vector store.
+
+    When provided to ``ContentUnderstandingContextProvider``, analyzed document
+    markdown is automatically uploaded to the specified vector store and the
+    given ``file_search`` tool is registered on the context. This enables
+    token-efficient RAG retrieval on follow-up turns for large documents.
+
+    The caller is responsible for creating and managing the vector store and
+    the ``file_search`` tool. Use :meth:`from_openai` or :meth:`from_foundry`
+    factory methods for convenience.
+
+    Args:
+        backend: A ``FileSearchBackend`` that handles file upload/delete
+            operations for the target vector store.
+        vector_store_id: The ID of a pre-existing vector store to upload to.
+        file_search_tool: A ``file_search`` tool object created via the LLM
+            client's ``get_file_search_tool()`` factory method. This is
+            registered on the context via ``extend_tools`` so the LLM can
+            retrieve uploaded content.
+    """
+
+    backend: FileSearchBackend
+    vector_store_id: str
+    file_search_tool: Any
+
+    @staticmethod
+    def from_openai(
+        client: Any,
+        *,
+        vector_store_id: str,
+        file_search_tool: Any,
+    ) -> FileSearchConfig:
+        """Create a config for OpenAI Responses API (``OpenAIChatClient``).
+
+        Args:
+            client: An ``AsyncOpenAI`` or ``AsyncAzureOpenAI`` client.
+            vector_store_id: The ID of the vector store to upload to.
+            file_search_tool: Tool from ``OpenAIChatClient.get_file_search_tool()``.
+        """
+        return FileSearchConfig(
+            backend=OpenAIFileSearchBackend(client),
+            vector_store_id=vector_store_id,
+            file_search_tool=file_search_tool,
+        )
+
+    @staticmethod
+    def from_foundry(
+        client: Any,
+        *,
+        vector_store_id: str,
+        file_search_tool: Any,
+    ) -> FileSearchConfig:
+        """Create a config for Azure AI Foundry (``FoundryChatClient``).
+
+        Args:
+            client: The OpenAI-compatible client from ``FoundryChatClient.client``.
+            vector_store_id: The ID of the vector store to upload to.
+            file_search_tool: Tool from ``FoundryChatClient.get_file_search_tool()``.
+        """
+        return FileSearchConfig(
+            backend=FoundryFileSearchBackend(client),
+            vector_store_id=vector_store_id,
+            file_search_tool=file_search_tool,
+        )