add file parsing and FAQ

alembic/versions/4fc61e385531_add_faq_and_chunk_config.py

@@ -0,0 +1,37 @@
+"""add chunk_config, FAQ tables and FAQ config columns
+
+Revision ID: 4fc61e385531
+Revises: 1432eea7c5b9
+Create Date: 2025-12-25 17:00:00.000000
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+from db.op.safe_add import safe_add_column
+
+
+# revision identifiers, used by Alembic.
+revision: str = '4fc61e385531'
+down_revision: Union[str, Sequence[str], None] = '1432eea7c5b9'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Upgrade schema."""
+    # 1. Add chunk_config column to pai_knowledgebase_file table
+    safe_add_column('pai_knowledgebase_file', sa.Column('chunk_config', sa.JSON(), nullable=True))
+    safe_add_column('pai_chatbot_model', sa.Column('enable_faq', sa.Boolean(), nullable=True, default=False))
+    safe_add_column('pai_chatbot_model', sa.Column('faq_config', sa.JSON(), nullable=True))
+
+
+def downgrade() -> None:
+    """Downgrade schema."""
+    # Remove chunk_config column from pai_knowledgebase_file table
+    op.drop_column('pai_knowledgebase_file', 'chunk_config')
+    op.drop_column('pai_chatbot_model', 'faq_config')
+    op.drop_column('pai_chatbot_model', 'enable_faq')
+

backend/agent/actor.py

@@ -14,6 +14,7 @@
 from common.chat.constants import MessageRole
 from opentelemetry import trace
 from utils.json_utils import parse_tool_arguments
+from agent.tool_utils import check_and_handle_return_direct
 
 
 @retry(stop=stop_after_attempt(3), wait=wait_fixed(1))
@@ -118,6 +119,19 @@ async def gen():
 
                         state.current_tool_call = None
                         observations += message_content + "\n\n"
+
+                        # Check if tool has return_direct=True, if so, return directly
+                        tool_obj = self.tool_fn_map[tool_name]
+                        return_chunk = check_and_handle_return_direct(
+                            tool_obj=tool_obj,
+                            tool_name=tool_name,
+                            tool_content=tool_content,
+                            tool_error=tool_error,
+                            agent_name=self.name,
+                        )
+                        if return_chunk:
+                            yield return_chunk
+                            return
                     act_prompt = self.build_prompt(state)
                     messages = [{"role": "system", "content": act_prompt}] + messages
 
@@ -212,6 +226,19 @@ async def gen():
                             )
                             observations += message_content + "\n\n"
 
+                            # Check if tool has return_direct=True, if so, return directly
+                            tool_obj = self.tool_fn_map[function_name]
+                            return_chunk = check_and_handle_return_direct(
+                                tool_obj=tool_obj,
+                                tool_name=function_name,
+                                tool_content=tool_content,
+                                tool_error=tool_error,
+                                agent_name=self.name,
+                            )
+                            if return_chunk:
+                                yield return_chunk
+                                return
+
                     # 超出步数保护
                     if react_step > self.max_steps:
                         logger.warning(f"Reached max recursion steps: {self.max_steps}")

backend/agent/actor_with_plan.py

@@ -10,6 +10,7 @@
 from extensions.trace.base import use_current_span
 from opentelemetry import trace
 from utils.json_utils import parse_tool_arguments
+from agent.tool_utils import check_and_handle_return_direct
 
 
 @retry(stop=stop_after_attempt(3), wait=wait_fixed(1))
@@ -157,6 +158,19 @@ async def gen():
                                         result=tool_content,
                                         error=tool_error
                                     )
+
+                                    # Check if tool has return_direct=True, if so, return directly
+                                    tool_obj = self.tool_fn_map[function_name]
+                                    return_chunk = check_and_handle_return_direct(
+                                        tool_obj=tool_obj,
+                                        tool_name=function_name,
+                                        tool_content=tool_content,
+                                        tool_error=tool_error,
+                                        agent_name=self.name,
+                                    )
+                                    if return_chunk:
+                                        yield return_chunk
+                                        return
                         else:
                             break
 

backend/agent/tool_utils.py

@@ -0,0 +1,60 @@
+"""Utility functions for handling tool calls and results."""
+
+import json
+from typing import Optional
+from llama_index.core.tools.function_tool import FunctionTool
+from common.llm.models import TextChunk
+from loguru import logger
+
+
+def check_and_handle_return_direct(
+    tool_obj: FunctionTool,
+    tool_name: str,
+    tool_content: Optional[str],
+    tool_error: Optional[str],
+    agent_name: str = "agent",
+) -> Optional[TextChunk]:
+    """
+    Check if a tool has return_direct=True and format the result accordingly.
+
+    Args:
+        tool_obj: The FunctionTool object
+        tool_name: Name of the tool
+        tool_content: Content returned by the tool (None if error)
+        tool_error: Error message if tool call failed (None if success)
+        agent_name: Name of the agent (for logging)
+
+    Returns:
+        TextChunk if return_direct=True, None otherwise
+    """
+    return_direct = getattr(tool_obj.metadata, 'return_direct', False)
+
+    if not return_direct:
+        return None
+
+    logger.info(f"[{agent_name}] Tool {tool_name} has return_direct=True, returning tool result directly.")
+
+    if tool_error:
+        return TextChunk(delta=f"工具调用失败: {tool_error}")
+
+    if not tool_content:
+        return TextChunk(delta="工具调用成功，但未返回内容。")
+
+    try:
+        result_data = json.loads(tool_content)
+        if isinstance(result_data, dict) and "result" in result_data:
+            # Format FAQ results or similar structured results
+            formatted_result = ""
+            for item in result_data.get("result", []):
+                if isinstance(item, dict):
+                    content = item.get("content", "")
+                    if content:
+                        formatted_result += content + "\n\n"
+            if formatted_result:
+                return TextChunk(delta=formatted_result.strip())
+            else:
+                return TextChunk(delta=tool_content)
+        else:
+            return TextChunk(delta=tool_content)
+    except (json.JSONDecodeError, Exception):
+        return TextChunk(delta=tool_content)