youngkingdom/dify
1
0
Fork 0
mirror of https://github.com/langgenius/dify.git synced 2026-04-29 23:18:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into feat/grouping-branching

Browse Source
This commit is contained in:
zhsama
2026-01-02 01:34:02 +08:00
parent 39d6383474 9b6b2f3195
commit bd338a9043
278 changed files with 3019 additions and 1015 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
api/core/agent/cot_agent_runner.py
View File

@ -22,6 +22,7 @@ from core.prompt.agent_history_prompt_transform import AgentHistoryPromptTransfo
from core.tools.__base.tool import Tool
from core.tools.entities.tool_entities import ToolInvokeMeta
from core.tools.tool_engine import ToolEngine
from core.workflow.nodes.agent.exc import AgentMaxIterationError
from models.model import Message
logger = logging.getLogger(__name__)
@ -165,6 +166,11 @@ class CotAgentRunner(BaseAgentRunner, ABC):
scratchpad.thought = scratchpad.thought.strip() or "I am thinking about how to help you"
self._agent_scratchpad.append(scratchpad)
# Check if max iteration is reached and model still wants to call tools
if iteration_step == max_iteration_steps and scratchpad.action:
if scratchpad.action.action_name.lower() != "final answer":
raise AgentMaxIterationError(app_config.agent.max_iteration)
# get llm usage
if "usage" in usage_dict:
if usage_dict["usage"] is not None:

5
api/core/agent/fc_agent_runner.py
View File

@ -25,6 +25,7 @@ from core.model_runtime.entities.message_entities import ImagePromptMessageConte
from core.prompt.agent_history_prompt_transform import AgentHistoryPromptTransform
from core.tools.entities.tool_entities import ToolInvokeMeta
from core.tools.tool_engine import ToolEngine
from core.workflow.nodes.agent.exc import AgentMaxIterationError
from models.model import Message
logger = logging.getLogger(__name__)
@ -222,6 +223,10 @@ class FunctionCallAgentRunner(BaseAgentRunner):
final_answer += response + "\n"
# Check if max iteration is reached and model still wants to call tools
if iteration_step == max_iteration_steps and tool_calls:
raise AgentMaxIterationError(app_config.agent.max_iteration)
# call tools
tool_responses = []
for tool_call_id, tool_call_name, tool_call_args in tool_calls:

44
api/core/rag/cleaner/clean_processor.py
View File

@ -27,26 +27,44 @@ class CleanProcessor:
pattern = r"([a-zA-Z0-9_.+-]+@[a-zA-Z0-9-]+\.[a-zA-Z0-9-.]+)"
text = re.sub(pattern, "", text)
# Remove URL but keep Markdown image URLs
# First, temporarily replace Markdown image URLs with a placeholder
markdown_image_pattern = r"!\[.*?\]\((https?://[^\s)]+)\)"
placeholders: list[str] = []
# Remove URL but keep Markdown image URLs and link URLs
# Replace the ENTIRE markdown link/image with a single placeholder to protect
# the link text (which might also be a URL) from being removed
markdown_link_pattern = r"\[([^\]]*)\]\((https?://[^)]+)\)"
markdown_image_pattern = r"!\[.*?\]\((https?://[^)]+)\)"
placeholders: list[tuple[str, str, str]] = [] # (type, text, url)
def replace_with_placeholder(match, placeholders=placeholders):
def replace_markdown_with_placeholder(match, placeholders=placeholders):
link_type = "link"
link_text = match.group(1)
url = match.group(2)
placeholder = f"__MARKDOWN_PLACEHOLDER_{len(placeholders)}__"
placeholders.append((link_type, link_text, url))
return placeholder
def replace_image_with_placeholder(match, placeholders=placeholders):
link_type = "image"
url = match.group(1)
placeholder = f"__MARKDOWN_IMAGE_URL_{len(placeholders)}__"
placeholders.append(url)
return f"![image]({placeholder})"
placeholder = f"__MARKDOWN_PLACEHOLDER_{len(placeholders)}__"
placeholders.append((link_type, "image", url))
return placeholder
text = re.sub(markdown_image_pattern, replace_with_placeholder, text)
# Protect markdown links first
text = re.sub(markdown_link_pattern, replace_markdown_with_placeholder, text)
# Then protect markdown images
text = re.sub(markdown_image_pattern, replace_image_with_placeholder, text)
# Now remove all remaining URLs
url_pattern = r"https?://[^\s)]+"
url_pattern = r"https?://\S+"
text = re.sub(url_pattern, "", text)
# Finally, restore the Markdown image URLs
for i, url in enumerate(placeholders):
text = text.replace(f"__MARKDOWN_IMAGE_URL_{i}__", url)
# Restore the Markdown links and images
for i, (link_type, text_or_alt, url) in enumerate(placeholders):
placeholder = f"__MARKDOWN_PLACEHOLDER_{i}__"
if link_type == "link":
text = text.replace(placeholder, f"[{text_or_alt}]({url})")
else: # image
text = text.replace(placeholder, f"![{text_or_alt}]({url})")
return text
def filter_string(self, text):

4
api/core/rag/extractor/extract_processor.py
View File

@ -112,7 +112,7 @@ class ExtractProcessor:
if file_extension in {".xlsx", ".xls"}:
extractor = ExcelExtractor(file_path)
elif file_extension == ".pdf":
extractor = PdfExtractor(file_path)
extractor = PdfExtractor(file_path, upload_file.tenant_id, upload_file.created_by)
elif file_extension in {".md", ".markdown", ".mdx"}:
extractor = (
UnstructuredMarkdownExtractor(file_path, unstructured_api_url, unstructured_api_key)
@ -148,7 +148,7 @@ class ExtractProcessor:
if file_extension in {".xlsx", ".xls"}:
extractor = ExcelExtractor(file_path)
elif file_extension == ".pdf":
extractor = PdfExtractor(file_path)
extractor = PdfExtractor(file_path, upload_file.tenant_id, upload_file.created_by)
elif file_extension in {".md", ".markdown", ".mdx"}:
extractor = MarkdownExtractor(file_path, autodetect_encoding=True)
elif file_extension in {".htm", ".html"}:

122
api/core/rag/extractor/pdf_extractor.py
View File

@ -1,25 +1,57 @@
"""Abstract interface for document loader implementations."""
import contextlib
import io
import logging
import uuid
from collections.abc import Iterator
import pypdfium2
import pypdfium2.raw as pdfium_c
from configs import dify_config
from core.rag.extractor.blob.blob import Blob
from core.rag.extractor.extractor_base import BaseExtractor
from core.rag.models.document import Document
from extensions.ext_database import db
from extensions.ext_storage import storage
from libs.datetime_utils import naive_utc_now
from models.enums import CreatorUserRole
from models.model import UploadFile
logger = logging.getLogger(__name__)
class PdfExtractor(BaseExtractor):
"""Load pdf files.
"""
PdfExtractor is used to extract text and images from PDF files.
Args:
file_path: Path to the file to load.
file_path: Path to the PDF file.
tenant_id: Workspace ID.
user_id: ID of the user performing the extraction.
file_cache_key: Optional cache key for the extracted text.
"""
def __init__(self, file_path: str, file_cache_key: str | None = None):
"""Initialize with file path."""
# Magic bytes for image format detection: (magic_bytes, extension, mime_type)
IMAGE_FORMATS = [
(b"\xff\xd8\xff", "jpg", "image/jpeg"),
(b"\x89PNG\r\n\x1a\n", "png", "image/png"),
(b"\x00\x00\x00\x0c\x6a\x50\x20\x20\x0d\x0a\x87\x0a", "jp2", "image/jp2"),
(b"GIF8", "gif", "image/gif"),
(b"BM", "bmp", "image/bmp"),
(b"II*\x00", "tiff", "image/tiff"),
(b"MM\x00*", "tiff", "image/tiff"),
(b"II+\x00", "tiff", "image/tiff"),
(b"MM\x00+", "tiff", "image/tiff"),
]
MAX_MAGIC_LEN = max(len(m) for m, _, _ in IMAGE_FORMATS)
def __init__(self, file_path: str, tenant_id: str, user_id: str, file_cache_key: str | None = None):
"""Initialize PdfExtractor."""
self._file_path = file_path
self._tenant_id = tenant_id
self._user_id = user_id
self._file_cache_key = file_cache_key
def extract(self) -> list[Document]:
@ -50,7 +82,6 @@ class PdfExtractor(BaseExtractor):
def parse(self, blob: Blob) -> Iterator[Document]:
"""Lazily parse the blob."""
import pypdfium2 # type: ignore
with blob.as_bytes_io() as file_path:
pdf_reader = pypdfium2.PdfDocument(file_path, autoclose=True)
@ -59,8 +90,87 @@ class PdfExtractor(BaseExtractor):
text_page = page.get_textpage()
content = text_page.get_text_range()
text_page.close()
image_content = self._extract_images(page)
if image_content:
content += "\n" + image_content
page.close()
metadata = {"source": blob.source, "page": page_number}
yield Document(page_content=content, metadata=metadata)
finally:
pdf_reader.close()
def _extract_images(self, page) -> str:
"""
Extract images from a PDF page, save them to storage and database,
and return markdown image links.
Args:
page: pypdfium2 page object.
Returns:
Markdown string containing links to the extracted images.
"""
image_content = []
upload_files = []
base_url = dify_config.INTERNAL_FILES_URL or dify_config.FILES_URL
try:
image_objects = page.get_objects(filter=(pdfium_c.FPDF_PAGEOBJ_IMAGE,))
for obj in image_objects:
try:
# Extract image bytes
img_byte_arr = io.BytesIO()
# Extract DCTDecode (JPEG) and JPXDecode (JPEG 2000) images directly
# Fallback to png for other formats
obj.extract(img_byte_arr, fb_format="png")
img_bytes = img_byte_arr.getvalue()
if not img_bytes:
continue
header = img_bytes[: self.MAX_MAGIC_LEN]
image_ext = None
mime_type = None
for magic, ext, mime in self.IMAGE_FORMATS:
if header.startswith(magic):
image_ext = ext
mime_type = mime
break
if not image_ext or not mime_type:
continue
file_uuid = str(uuid.uuid4())
file_key = "image_files/" + self._tenant_id + "/" + file_uuid + "." + image_ext
storage.save(file_key, img_bytes)
# save file to db
upload_file = UploadFile(
tenant_id=self._tenant_id,
storage_type=dify_config.STORAGE_TYPE,
key=file_key,
name=file_key,
size=len(img_bytes),
extension=image_ext,
mime_type=mime_type,
created_by=self._user_id,
created_by_role=CreatorUserRole.ACCOUNT,
created_at=naive_utc_now(),
used=True,
used_by=self._user_id,
used_at=naive_utc_now(),
)
upload_files.append(upload_file)
image_content.append(f"![image]({base_url}/files/{upload_file.id}/file-preview)")
except Exception as e:
logger.warning("Failed to extract image from PDF: %s", e)
continue
except Exception as e:
logger.warning("Failed to get objects from PDF page: %s", e)
if upload_files:
db.session.add_all(upload_files)
db.session.commit()
return "\n".join(image_content)

2
api/core/tools/utils/parser.py
View File

@ -378,7 +378,7 @@ class ApiBasedToolSchemaParser:
@staticmethod
def auto_parse_to_tool_bundle(
content: str, extra_info: dict | None = None, warning: dict | None = None
) -> tuple[list[ApiToolBundle], str]:
) -> tuple[list[ApiToolBundle], ApiProviderSchemaType]:
"""
auto parse to tool bundle

6
api/core/tools/utils/text_processing_utils.py
View File

@ -4,6 +4,7 @@ import re
def remove_leading_symbols(text: str) -> str:
"""
Remove leading punctuation or symbols from the given text.
Preserves markdown links like [text](url) at the start.
Args:
text (str): The input text to process.
@ -11,6 +12,11 @@ def remove_leading_symbols(text: str) -> str:
Returns:
str: The text with leading punctuation or symbols removed.
"""
# Check if text starts with a markdown link - preserve it
markdown_link_pattern = r"^\[([^\]]+)\]\((https?://[^)]+)\)"
if re.match(markdown_link_pattern, text):
return text
# Match Unicode ranges for punctuation and symbols
# FIXME this pattern is confused quick fix for #11868 maybe refactor it later
pattern = r'^[\[\]\u2000-\u2025\u2027-\u206F\u2E00-\u2E7F\u3000-\u300F\u3011-\u303F"#$%&\'()*+,./:;<=>?@^_`~]+'

1
api/core/workflow/graph_engine/graph_traversal/skip_propagator.py
View File

@ -60,6 +60,7 @@ class SkipPropagator:
if edge_states["has_taken"]:
# Enqueue node
self._state_manager.enqueue_node(downstream_node_id)
self._state_manager.start_execution(downstream_node_id)
return
# All edges are skipped, propagate skip to this node

11
api/core/workflow/nodes/agent/exc.py
View File

@ -119,3 +119,14 @@ class AgentVariableTypeError(AgentNodeError):
self.expected_type = expected_type
self.actual_type = actual_type
super().__init__(message)
class AgentMaxIterationError(AgentNodeError):
"""Exception raised when the agent exceeds the maximum iteration limit."""
def __init__(self, max_iteration: int):
self.max_iteration = max_iteration
super().__init__(
f"Agent exceeded the maximum iteration limit of {max_iteration}. "
f"The agent was unable to complete the task within the allowed number of iterations."
)