| |
| |
| |
| |
| @@ -354,8 +354,13 @@ async def run_agent( |
| print(f"Turn {turn}") |
| turn_trajectory: dict = {"turn": turn, "messages": list(messages)} |
| |
| + # Agent (orchestrator) endpoint may differ from the vision endpoint: |
| + # vision tools use llm_base_url (e.g. local vLLM); the agent uses agent_base_url |
| + # (e.g. OpenRouter/deepseek) when set, falling back to the llm_* pair. |
| + _agent_base = settings.agent_base_url or settings.llm_base_url |
| + _agent_key = settings.agent_api_key or settings.llm_api_key |
| resp = await _send_request( |
| - settings.llm_base_url, settings.llm_api_key, model, messages |
| + _agent_base, _agent_key, model, messages |
| ) |
| if "content" not in resp: |
| print(f"No content in response: {resp}") |
| |
| |
| |
| |
| @@ -28,6 +28,22 @@ def video_to_data_url(path: str, mime="video/mp4") -> str: |
| return f"data:{mime};base64,{b64}" |
| |
| |
| +def _clip_to_frame_urls(mp4: str, fps: float = 0.5, cap: int = 28, dim: int = 768) -> list: |
| + """Sample chronological frames from a clip mp4 as base64 image data-URLs. |
| + Used for image-only vision models (e.g. qwen) that reject video_url input.""" |
| + import subprocess, tempfile, glob as _glob |
| + d = tempfile.mkdtemp(prefix="clipfr_") |
| + subprocess.run( |
| + ["ffmpeg", "-nostdin", "-loglevel", "error", "-i", mp4, "-vf", |
| + f"fps={fps},scale='if(gt(iw,ih),{dim},-2)':'if(gt(iw,ih),-2,{dim})'", "-q:v", "4", |
| + os.path.join(d, "f%04d.jpg")], check=False) |
| + out = [] |
| + for p in sorted(_glob.glob(os.path.join(d, "*.jpg")))[:cap]: |
| + with open(p, "rb") as f: |
| + out.append("data:image/jpeg;base64," + base64.b64encode(f.read()).decode()) |
| + return out |
| + |
| + |
| def construct_payload(clips: Optional[List[Clip]] = None, frames: Optional[List[Frame]] = None): |
| payload = [] |
| if clips is None and frames is None: |
| @@ -38,28 +54,26 @@ def construct_payload(clips: Optional[List[Clip]] = None, frames: Optional[List[ |
| is_clip_url = clip.clip_url is not None |
| is_clip_file_path = clip.clip_file_path is not None and os.path.exists(clip.clip_file_path) |
| |
| - if is_clip_url: |
| - url = clip.clip_url |
| - elif is_clip_file_path: |
| - url = video_to_data_url(clip.clip_file_path) |
| - else: |
| - raise ValueError(f"Clip {clip.id} has no valid url or file path") |
| - |
| - |
| timestamp = f"Timestamp: {clip.start_time} seconds to {clip.end_time} seconds" |
| - |
| - payload.extend( |
| - [ |
| - { |
| - "type": "text", |
| - "text": f"Clip ID: {clip.id}\n{timestamp}", |
| - }, |
| - { |
| - "type": "video_url", |
| - "video_url": {"url": url}, |
| - }, |
| - ] |
| - ) |
| + image_only = "gemini" not in (settings.llm_model or "").lower() |
| + |
| + if is_clip_file_path and image_only: |
| + # image-only vision model (e.g. qwen3.6-27b): send sampled frames, not video_url |
| + payload.append({"type": "text", |
| + "text": f"Clip ID: {clip.id}\n{timestamp}\n(chronological sampled frames)"}) |
| + for u in _clip_to_frame_urls(clip.clip_file_path): |
| + payload.append({"type": "image_url", "image_url": {"url": u}}) |
| + else: |
| + if is_clip_url: |
| + url = clip.clip_url |
| + elif is_clip_file_path: |
| + url = video_to_data_url(clip.clip_file_path) |
| + else: |
| + raise ValueError(f"Clip {clip.id} has no valid url or file path") |
| + payload.extend([ |
| + {"type": "text", "text": f"Clip ID: {clip.id}\n{timestamp}"}, |
| + {"type": "video_url", "video_url": {"url": url}}, |
| + ]) |
| |
| if frames is not None: |
| start_time = frames[0].timestamp |
|
|
