Merge branch 'main' into full-claude-35-support

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "vision-agent"
-version = "0.2.133"
+version = "0.2.135"
 description = "Toolset for Vision Agent"
 authors = ["Landing AI <[email protected]>"]
 readme = "README.md"

vision_agent/agent/vision_agent.py

@@ -14,7 +14,7 @@
  VA_CODE,
 )
 from vision_agent.lmm import LMM, AnthropicLMM, Message, OpenAILMM
-from vision_agent.tools import META_TOOL_DOCSTRING, load_image, save_image
+from vision_agent.tools import META_TOOL_DOCSTRING, extract_frames, load_image, save_image
 from vision_agent.tools.meta_tools import (
  Artifacts,
  check_and_load_image,
@@ -235,9 +235,20 @@ def chat_with_code(
  for media in chat_i["media"]:
  if type(media) is str and media.startswith(("http", "https")):
  # TODO: Ideally we should not call VA.tools here, we should come to revisit how to better support remote image later
- file_path = Path(media).name
- ndarray = load_image(media)
- save_image(ndarray, file_path)
+ file_path = str(
+ Path(self.local_artifacts_path).parent
+ / Path(media).name
+ )
+ if file_path.lower().endswith(
+ ".mp4"
+ ) or file_path.lower().endswith(".mov"):
+ video_frames = extract_frames(media)
+ save_video(
+ [frame for frame, _ in video_frames], file_path
+ )
+ else:
+ ndarray = load_image(media)
+ save_image(ndarray, file_path)
  media = file_path
  else:
  media = cast(str, media)

vision_agent/tools/meta_tools.py

@@ -53,25 +53,27 @@ def redisplay_results(execution: Execution) -> None:
  """
  for result in execution.results:
  if result.text is not None:
- display({MimeType.TEXT_PLAIN: result.text})
+ display({MimeType.TEXT_PLAIN: result.text}, raw=True)
  if result.html is not None:
- display({MimeType.TEXT_HTML: result.html})
+ display({MimeType.TEXT_HTML: result.html}, raw=True)
  if result.markdown is not None:
- display({MimeType.TEXT_MARKDOWN: result.markdown})
+ display({MimeType.TEXT_MARKDOWN: result.markdown}, raw=True)
  if result.svg is not None:
- display({MimeType.IMAGE_SVG: result.svg})
+ display({MimeType.IMAGE_SVG: result.svg}, raw=True)
  if result.png is not None:
- display({MimeType.IMAGE_PNG: result.png})
+ display({MimeType.IMAGE_PNG: result.png}, raw=True)
  if result.jpeg is not None:
- display({MimeType.IMAGE_JPEG: result.jpeg})
+ display({MimeType.IMAGE_JPEG: result.jpeg}, raw=True)
  if result.mp4 is not None:
- display({MimeType.VIDEO_MP4_B64: result.mp4})
+ display({MimeType.VIDEO_MP4_B64: result.mp4}, raw=True)
  if result.latex is not None:
- display({MimeType.TEXT_LATEX: result.latex})
+ display({MimeType.TEXT_LATEX: result.latex}, raw=True)
  if result.json is not None:
- display({MimeType.APPLICATION_JSON: result.json})
+ display({MimeType.APPLICATION_JSON: result.json}, raw=True)
+ if result.artifact_name is not None:
+ display({MimeType.TEXT_ARTIFACT_NAME: result.artifact_name}, raw=True)
  if result.extra is not None:
- display(result.extra)
+ display(result.extra, raw=True)
 
 
 class Artifacts:
@@ -208,7 +210,7 @@ def create_code_artifact(artifacts: Artifacts, name: str) -> str:
  return_str = f"[Artifact {name} created]"
  print(return_str)
 
- display({MimeType.APPLICATION_JSON: {"last_artifact": name}})
+ display({MimeType.TEXT_ARTIFACT_NAME: name}, raw=True)
  return return_str
 
 
@@ -292,7 +294,7 @@ def edit_code_artifact(
 
  artifacts[name] = "".join(edited_lines)
 
- display({MimeType.APPLICATION_JSON: {"last_artifact": name}})
+ display({MimeType.TEXT_ARTIFACT_NAME: name}, raw=True)
  return open_code_artifact(artifacts, name, cur_line)
 
 
@@ -348,7 +350,7 @@ def detect_dogs(image_path: str):
  code_lines = code.splitlines(keepends=True)
  total_lines = len(code_lines)
 
- display({MimeType.APPLICATION_JSON: {"last_artifact": name}})
+ display({MimeType.TEXT_ARTIFACT_NAME: name}, raw=True)
  return view_lines(code_lines, 0, total_lines, name, total_lines)
 
 
@@ -413,7 +415,7 @@ def detect_dogs(image_path: str):
  code_lines = code.splitlines(keepends=True)
  total_lines = len(code_lines)
 
- display({MimeType.APPLICATION_JSON: {"last_artifact": name}})
+ display({MimeType.TEXT_ARTIFACT_NAME: name}, raw=True)
  return view_lines(code_lines, 0, total_lines, name, total_lines)
 
 
@@ -427,6 +429,7 @@ def write_media_artifact(artifacts: Artifacts, local_path: str) -> str:
  with open(local_path, "rb") as f:
  media = f.read()
  artifacts[Path(local_path).name] = media
+ display({MimeType.TEXT_ARTIFACT_NAME: Path(local_path).name}, raw=True)
  return f"[Media {Path(local_path).name} saved]"
 
 
@@ -623,6 +626,8 @@ def use_object_detection_fine_tuning(
 
  diff = get_diff_with_prompts(name, code, new_code)
  print(diff)
+
+ display({MimeType.TEXT_ARTIFACT_NAME: name}, raw=True)
  return diff
 
 

vision_agent/utils/execute.py

@@ -56,6 +56,7 @@ class MimeType(str, Enum):
  TEXT_LATEX = "text/latex"
  APPLICATION_JSON = "application/json"
  APPLICATION_JAVASCRIPT = "application/javascript"
+ TEXT_ARTIFACT_NAME = "text/artifact/name"
 
 
 class FileSerializer:
@@ -103,6 +104,7 @@ class Result:
  latex: Optional[str] = None
  json: Optional[Dict[str, Any]] = None
  javascript: Optional[str] = None
+ artifact_name: Optional[str] = None
  extra: Optional[Dict[str, Any]] = None
  "Extra data that can be included. Not part of the standard types."
 
@@ -127,6 +129,7 @@ def __init__(self, is_main_result: bool, data: Dict[str, Any]):
  self.latex = data.pop(MimeType.TEXT_LATEX, None)
  self.json = data.pop(MimeType.APPLICATION_JSON, None)
  self.javascript = data.pop(MimeType.APPLICATION_JAVASCRIPT, None)
+ self.artifact_name = data.pop(MimeType.TEXT_ARTIFACT_NAME, None)
  self.extra = data
  # Only keeping the PNG representation if both PNG and JPEG are present
  if self.png and self.jpeg:
@@ -204,6 +207,8 @@ def formats(self) -> Iterable[str]:
  formats.append("javascript")
  if self.mp4:
  formats.append("mp4")
+ if self.artifact_name:
+ formats.append("artifact_name")
  if self.extra:
  formats.extend(iter(self.extra))
  return formats
@@ -691,8 +696,9 @@ def new_instance(
  if not code_sandbox_runtime:
  code_sandbox_runtime = os.getenv("CODE_SANDBOX_RUNTIME", "local")
  if code_sandbox_runtime == "e2b":
+ envs = _get_e2b_env()
  instance: CodeInterpreter = E2BCodeInterpreter(
- timeout=_SESSION_TIMEOUT, remote_path=remote_path
+ timeout=_SESSION_TIMEOUT, remote_path=remote_path, envs=envs
  )
  elif code_sandbox_runtime == "local":
  instance = LocalCodeInterpreter(
@@ -705,6 +711,20 @@ def new_instance(
  return instance
 
 
+def _get_e2b_env() -> Union[Dict[str, str], None]:
+ openai_api_key = os.getenv("OPENAI_API_KEY", "")
+ anthropic_api_key = os.getenv("ANTHROPIC_API_KEY", "")
+ if openai_api_key or anthropic_api_key:
+ envs = {}
+ if openai_api_key:
+ envs["OPENAI_API_KEY"] = openai_api_key
+ if anthropic_api_key:
+ envs["ANTHROPIC_API_KEY"] = anthropic_api_key
+ else:
+ envs = None
+ return envs
+
+
 def _parse_local_code_interpreter_outputs(outputs: List[Dict[str, Any]]) -> Execution:
  """Parse notebook cell outputs to Execution object. Output types:
  https://nbformat.readthedocs.io/en/latest/format_description.html#code-cell-outputs