xusijie
Clean branch for HF push
06ba7ea
import asyncio
import os
import math
import base64
from io import BytesIO
from typing import Any, Dict, List, Tuple
from urllib.parse import urlparse
from PIL import Image
from moviepy.video.io.VideoFileClip import VideoFileClip
from langchain_core.messages import SystemMessage, HumanMessage, AIMessage
from mcp.types import CreateMessageRequestParams, CreateMessageResult, TextContent
# -----------------------------
# Configurable parameters: Control multimodal input size
# -----------------------------
DEFAULT_RESIZE_EDGE = 600
DEFAULT_JPEG_QUALITY = 80
DEFAULT_MIN_FRAMES = 2
DEFAULT_MAX_FRAMES = 6
DEFAULT_FRAMES_PER_SEC = 3.0
GLOBAL_MAX_IMAGE_BLOCKS = 48 # Maximum total images allowed (video frames + images) to prevent payload overflow
IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".webp", ".bmp", ".gif", ".tiff"}
VIDEO_EXTS = {".mp4", ".mov", ".mkv", ".avi", ".webm", ".m4v"}
def _is_data_url(u: str) -> bool:
return isinstance(u, str) and u.startswith("data:")
def _is_http_url(u: str) -> bool:
return isinstance(u, str) and (u.startswith("http://") or u.startswith("https://"))
def _strip_file_scheme(u: str) -> str:
if not isinstance(u, str):
return str(u)
if u.startswith("file://"):
parsed = urlparse(u)
return parsed.path
return u
def _guess_ext(path_or_url: str) -> str:
try:
p = urlparse(path_or_url).path if _is_http_url(path_or_url) else path_or_url
return os.path.splitext(p)[1].lower()
except Exception:
return ""
def _resize_long_edge(img: Image.Image, long_edge: int) -> Image.Image:
if long_edge <= 0:
return img
w, h = img.size
le = max(w, h)
if le <= long_edge:
return img
scale = long_edge / float(le)
nw = max(1, int(round(w * scale)))
nh = max(1, int(round(h * scale)))
return img.resize((nw, nh), Image.LANCZOS)
def _pil_to_data_url(img: Image.Image, resize_edge: int, jpeg_quality: int) -> str:
img = img.convert("RGB")
img = _resize_long_edge(img, resize_edge)
buf = BytesIO()
img.save(buf, format="JPEG", quality=jpeg_quality, optimize=True)
b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
return f"data:image/jpeg;base64,{b64}"
def _image_path_to_data_url(path: str, resize_edge: int, jpeg_quality: int) -> str:
img = Image.open(path)
return _pil_to_data_url(img, resize_edge, jpeg_quality)
def _choose_num_frames(duration_sec: float, min_frames: int, max_frames: int, frames_per_sec: float) -> int:
duration_sec = max(0.0, float(duration_sec))
n = int(math.ceil(duration_sec * frames_per_sec))
n = max(min_frames, n)
n = min(max_frames, n)
return n
def _sample_video_segment_to_data_urls(
video_path: str,
in_sec: float,
out_sec: float,
resize_edge: int,
jpeg_quality: int,
min_frames: int,
max_frames: int,
frames_per_sec: float,
) -> List[Tuple[float, str]]:
"""
Sample frames only from the [in_sec, out_sec] segment. Returns (rel_t_from_in, data_url)
"""
in_sec = float(in_sec)
out_sec = float(out_sec)
clip = VideoFileClip(video_path, audio=False)
try:
vdur = float(clip.duration or 0.0)
# If duration is unavailable, conservatively sample one frame to avoid out_sec exceeding bounds
if vdur <= 0:
t = max(0.0, in_sec)
frame = clip.get_frame(t)
img = Image.fromarray(frame)
return [(0.0, _pil_to_data_url(img, resize_edge, jpeg_quality))]
# Clamp to valid range
in_sec = max(0.0, min(in_sec, vdur))
out_sec = max(0.0, min(out_sec, vdur))
# If still invalid, fallback to one frame at in_sec
if out_sec <= in_sec:
frame = clip.get_frame(in_sec)
img = Image.fromarray(frame)
return [(0.0, _pil_to_data_url(img, resize_edge, jpeg_quality))]
seg_dur = out_sec - in_sec
n = _choose_num_frames(seg_dur, min_frames, max_frames, frames_per_sec)
# Sample at bucket centers to avoid boundary frames
times = [((i + 0.5) / n) * seg_dur for i in range(n)]
out: List[Tuple[float, str]] = []
for rel_t in times:
abs_t = in_sec + rel_t
frame = clip.get_frame(abs_t)
img = Image.fromarray(frame)
out.append((rel_t, _pil_to_data_url(img, resize_edge, jpeg_quality)))
return out
finally:
clip.close()
def _extract_text_from_mcp_content(content: Any) -> str:
if content is None:
return ""
blocks = content if isinstance(content, list) else [content]
texts: List[str] = []
for b in blocks:
if getattr(b, "type", None) == "text":
texts.append(getattr(b, "text", "") or "")
return "\n".join([t for t in (x.strip() for x in texts) if t])
def _extract_text_from_lc_response(resp: Any) -> str:
content = getattr(resp, "content", None)
if isinstance(content, str):
return content.strip()
if isinstance(content, list):
texts = []
for blk in content:
if isinstance(blk, dict) and blk.get("type") == "text":
texts.append(str(blk.get("text", "")).strip())
return "\n".join([t for t in texts if t]).strip()
return str(resp).strip()
def _normalize_media_items(media_inputs: List[Any]) -> List[Dict[str, Any]]:
"""
Supports three input formats:
1) "path/to/video.mp4"
2) {"url"/"path": "...", "in_sec": 1.2, "out_sec": 3.4}
3) ("path/to/video.mp4", 1.2, 3.4) # optional
Output normalized to: {"url": "...", "in_sec": optional, "out_sec": optional}
"""
out = []
for item in media_inputs or []:
if isinstance(item, str):
out.append({"url": item})
continue
if isinstance(item, (list, tuple)) and len(item) >= 1:
d = {"url": item[0]}
if len(item) >= 2:
d["in_sec"] = item[1]
if len(item) >= 3:
d["out_sec"] = item[2]
out.append(d)
continue
if isinstance(item, dict):
url = item.get("url") or item.get("path") or item.get("media")
if not url:
continue
d = {"url": url}
if "in_sec" in item:
d["in_sec"] = item.get("in_sec")
if "out_sec" in item:
d["out_sec"] = item.get("out_sec")
out.append(d)
continue
return out
def _build_media_blocks(
media_inputs: List[Any],
resize_edge: int,
jpeg_quality: int,
min_frames: int,
max_frames: int,
frames_per_sec: float,
global_max_images: int,
) -> List[Dict[str, Any]]:
"""
Convert media_inputs to OpenAI-compatible multimodal blocks.
Videos: Sample frames by segment (if in_sec/out_sec provided) or entire video (legacy format)
"""
blocks: List[Dict[str, Any]] = []
img_count = 0
items = _normalize_media_items(media_inputs)
for idx, mi in enumerate(items):
if img_count >= global_max_images:
break
raw_url = _strip_file_scheme(str(mi.get("url")))
ext = _guess_ext(raw_url)
in_sec = mi.get("in_sec")
out_sec = mi.get("out_sec")
has_segment = (in_sec is not None and out_sec is not None)
# 1) Data URL (image) - pass through directly
if _is_data_url(raw_url):
blocks.append({"type": "text", "text": f"Media {idx+1}: (data url image)"})
blocks.append({"type": "image_url", "image_url": {"url": raw_url}})
img_count += 1
continue
# 2) Remote URL: Images can be passed through; remote videos cannot be sampled locally (provide notice)
if _is_http_url(raw_url):
if ext in VIDEO_EXTS:
seg_info = f" segment [{in_sec},{out_sec}]s" if has_segment else ""
blocks.append({"type": "text", "text": f"Media {idx+1}: remote video url{seg_info} (cannot sample frames locally): {raw_url}"})
continue
blocks.append({"type": "text", "text": f"Media {idx+1}: {raw_url}"})
blocks.append({"type": "image_url", "image_url": {"url": raw_url}})
img_count += 1
continue
# 3) Local path
path = raw_url
if not os.path.exists(path):
blocks.append({"type": "text", "text": f"Media {idx+1}: (missing file) {path}"})
continue
# Image
if ext in IMAGE_EXTS:
data_url = _image_path_to_data_url(path, resize_edge, jpeg_quality)
blocks.append({"type": "text", "text": f"Media {idx+1}: image file {os.path.basename(path)}"})
blocks.append({"type": "image_url", "image_url": {"url": data_url}})
img_count += 1
continue
# Video (supports segmented sampling)
if ext in VIDEO_EXTS:
if has_segment:
in_s = float(in_sec)
out_s = float(out_sec)
else:
# Legacy format: entire video. Use [0, +inf], internally clamped to duration
in_s = 0.0
out_s = 1e12
frames = _sample_video_segment_to_data_urls(
video_path=path,
in_sec=in_s,
out_sec=out_s,
resize_edge=resize_edge,
jpeg_quality=jpeg_quality,
min_frames=min_frames,
max_frames=max_frames,
frames_per_sec=frames_per_sec,
)
if has_segment:
blocks.append({"type": "text", "text": f"Media {idx+1}: video segment {os.path.basename(path)} [{in_s:.2f}s, {out_s:.2f}s] (sampled frames in time order)"})
else:
blocks.append({"type": "text", "text": f"Media {idx+1}: video file {os.path.basename(path)} (sampled frames in time order)"})
for fi, (rel_t, data_url) in enumerate(frames):
if img_count >= global_max_images:
break
blocks.append({"type": "text", "text": f"Frame {fi+1}/{len(frames)} (t≈{rel_t:.2f}s from segment start)"})
blocks.append({"type": "image_url", "image_url": {"url": data_url}})
img_count += 1
continue
blocks.append({"type": "text", "text": f"Media {idx+1}: unsupported file type: {path}"})
return blocks
def make_sampling_callback(
llm,
vlm,
*,
resize_edge: int = DEFAULT_RESIZE_EDGE,
jpeg_quality: int = DEFAULT_JPEG_QUALITY,
min_frames: int = DEFAULT_MIN_FRAMES,
max_frames: int = DEFAULT_MAX_FRAMES,
frames_per_sec: float = DEFAULT_FRAMES_PER_SEC,
global_max_images: int = GLOBAL_MAX_IMAGE_BLOCKS,
):
"""
Callback for MCP server sampling requests within tools:
- Reads metadata.media_urls (supports in_sec/out_sec)
- Samples frames and constructs LangChain multimodal messages
- Selects llm/vlm based on presence of media input
"""
async def sampling_callback(context, params: CreateMessageRequestParams) -> CreateMessageResult:
try:
# 1. System prompt
system_prompt = getattr(params, "systemPrompt", None) or ""
# 2. MCP messages (multi-turn) -> LangChain
mcp_messages = getattr(params, "messages", []) or []
lc_messages: List[Any] = []
if system_prompt:
lc_messages.append(SystemMessage(content=system_prompt))
# 3. Metadata: Extract media_urls and top_p
metadata = getattr(params, "metadata", None) or {}
media_inputs = list(metadata.get("media", []) or [])
top_p: float = float(metadata.get("top_p", 0.9))
temperature: float = float(getattr(params, "temperature", None) or 0.6)
max_tokens: int = int(getattr(params, "maxTokens", 4096) or 4096)
# 4. Route to appropriate model
use_multimodal = bool(media_inputs)
model = vlm if use_multimodal else llm
if model is None:
model = vlm or llm
# 5. Build media blocks (including video segment sampling) - run in thread to avoid blocking event loop
media_blocks: List[Dict[str, Any]] = []
if use_multimodal:
media_blocks = await asyncio.to_thread(
_build_media_blocks,
media_inputs,
resize_edge,
jpeg_quality,
min_frames,
max_frames,
frames_per_sec,
global_max_images,
)
# 6. Attach media to "last user message"
user_indices = [i for i, m in enumerate(mcp_messages) if getattr(m, "role", "") == "user"]
last_user_idx = user_indices[-1] if user_indices else None
if not mcp_messages:
# No messages - create a user message
content_blocks = [{"type": "text", "text": ""}]
if media_blocks:
content_blocks.extend(media_blocks)
lc_messages.append(HumanMessage(content=content_blocks if media_blocks else ""))
else:
for i, m in enumerate(mcp_messages):
role = getattr(m, "role", "") or "user"
text = _extract_text_from_mcp_content(getattr(m, "content", None))
if role == "assistant":
lc_messages.append(AIMessage(content=text))
continue
if role == "user":
if last_user_idx is not None and i == last_user_idx and media_blocks:
content_blocks = [{"type": "text", "text": text}]
content_blocks.extend(media_blocks)
lc_messages.append(HumanMessage(content=content_blocks))
else:
lc_messages.append(HumanMessage(content=text))
continue
lc_messages.append(HumanMessage(content=text))
# 7. Invoke selected model
bound = model
model_name = getattr(model, "model", None) or getattr(model, "model_name", None) or "unknown"
try:
bound = bound.bind(temperature=temperature, max_tokens=max_tokens, top_p=top_p)
except Exception:
bound = bound.bind(temperature=temperature, max_tokens=max_tokens)
try:
if hasattr(bound, "ainvoke"):
resp = await bound.ainvoke(lc_messages)
else:
resp = await asyncio.to_thread(bound.invoke, lc_messages)
except TypeError:
# Edge case: some wrappers don't accept max_tokens/top_p
bound2 = model.bind(temperature=temperature)
if hasattr(bound2, "ainvoke"):
resp = await bound2.ainvoke(lc_messages)
else:
resp = await asyncio.to_thread(bound2.invoke, lc_messages)
text_out = _extract_text_from_lc_response(resp)
return CreateMessageResult(
content=TextContent(type="text", text=text_out),
model=str(model_name),
role="assistant",
stopReason="endTurn",
)
except Exception as e:
return CreateMessageResult(
content=TextContent(type="text", text=f"{type(e)}: {e}"),
model=str(model_name),
role="assistant",
stopReason="error",
)
return sampling_callback