.

timesfm_backend.py

@@ -49,6 +49,54 @@ def _as_1d_float_tensor(series: List[float], device: str, dtype: torch.dtype) ->
     t = torch.tensor(series, dtype=torch.float32)  # keep input parse stable
     return t.to(device=device, dtype=dtype)
 
+
+    async def stream(self, request: Dict[str, Any]):
+        """
+        Chat-compatible shim: runs forecast() and yields a single OpenAI-style
+        chat.completion.chunk where the assistant's content is a compact JSON
+        payload containing the forecast result.
+        """
+        rid = f"chatcmpl-timesfm-{int(time.time())}"
+        now = int(time.time())
+
+        # map chat-ish payloads to our forecast inputs
+        # accept both direct {series,horizon,freq,...} and nested {data:{...}}
+        payload = request or {}
+        if "data" in payload and isinstance(payload["data"], dict):
+            payload = payload["data"]
+
+        # allow callers to stick series/fc config under "timeseries" too
+        if "timeseries" in payload and isinstance(payload["timeseries"], dict):
+            payload = {**payload, **payload["timeseries"]}
+
+        # run the forecast
+        result = await self.forecast(payload)
+
+        # minimal content: JSON string so your .NET backend can parse easily
+        import json as _json
+        content = _json.dumps({
+            "model": result.get("model"),
+            "horizon": result.get("horizon"),
+            "freq": result.get("freq"),
+            "forecast": result.get("forecast"),
+            "note": result.get("note", None),
+            "backend": "timesfm"
+        }, separators=(",", ":"), ensure_ascii=False)
+
+        yield {
+            "id": rid,
+            "object": "chat.completion.chunk",
+            "created": now,
+            "model": result.get("model"),
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {"role": "assistant", "content": content},
+                    "finish_reason": "stop",
+                }
+            ],
+        }
+
 # --------------------------------------------------------------------------------------
 # Fallback forecaster (naive)
 # --------------------------------------------------------------------------------------