Extract Hub repo artifacts from response text

backend/openai_compat.py

@@ -32,6 +32,19 @@ TERMINAL_RESPONSE_STATUSES = {"completed", "failed", "cancelled"}
 
 # HF job URLs look like https://huggingface.co/jobs/<namespace>/<job_id>
 _JOB_URL_ID_RE = re.compile(r"/jobs/[^/?#]+/([^/?#]+)/?(?:[?#]|$)")
+_HF_DATASET_URL_RE = re.compile(
+    r"https://huggingface\.co/datasets/"
+    r"([A-Za-z0-9][\w.-]*/[A-Za-z0-9][\w.-]*)(?:[/?#]|$)"
+)
+_HF_SPACE_URL_RE = re.compile(
+    r"https://huggingface\.co/spaces/"
+    r"([A-Za-z0-9][\w.-]*/[A-Za-z0-9][\w.-]*)(?:[/?#]|$)"
+)
+_HF_MODEL_URL_RE = re.compile(
+    r"https://huggingface\.co/"
+    r"(?!(?:datasets|spaces|jobs|papers|collections)/)"
+    r"([A-Za-z0-9][\w.-]*/[A-Za-z0-9][\w.-]*)(?:[/?#]|$)"
+)
 
 
 class V1APIError(Exception):
@@ -85,6 +98,29 @@ def _job_id_from_url(url: str) -> str | None:
     return match.group(1) if match else None
 
 
+def _repo_artifacts_from_text(text: str | None) -> list[dict[str, Any]]:
+    """Extract Hub repo links from assistant/tool text as best-effort artifacts."""
+    if not isinstance(text, str) or "https://huggingface.co/" not in text:
+        return []
+
+    artifacts: list[dict[str, Any]] = []
+    for repo_type, pattern, prefix in (
+        ("dataset", _HF_DATASET_URL_RE, "datasets/"),
+        ("space", _HF_SPACE_URL_RE, "spaces/"),
+        ("model", _HF_MODEL_URL_RE, ""),
+    ):
+        for match in pattern.finditer(text):
+            repo_id = match.group(1)
+            artifacts.append(
+                {
+                    "type": repo_type,
+                    "repo_id": repo_id,
+                    "url": f"https://huggingface.co/{prefix}{repo_id}",
+                }
+            )
+    return merge_artifacts([], artifacts)
+
+
 def artifact_key(artifact: dict[str, Any]) -> str:
     ident = (
         artifact.get("id")
@@ -136,6 +172,8 @@ def artifacts_from_event(event_type: str, data: dict[str, Any]) -> list[dict[str
                     "url": f"https://huggingface.co/{prefix}{repo_id}",
                 }
             )
+    for key in ("content", "final_response", "output"):
+        artifacts.extend(_repo_artifacts_from_text(data.get(key)))
     return artifacts