markdown-flow 0.2.15__py3-none-any.whl → 0.2.17__py3-none-any.whl

markdown_flow/core.py

@@ -6,7 +6,7 @@ Refactored MarkdownFlow class with built-in LLM processing capabilities and unif
 
 import json
 import re
-from collections.abc import Generator
+from collections.abc import AsyncGenerator
 from copy import copy
 from typing import Any
 
@@ -28,6 +28,7 @@ from .constants import (
     INTERACTION_PATTERN_SPLIT,
     INTERACTION_RENDER_INSTRUCTIONS,
     LLM_PROVIDER_REQUIRED_ERROR,
+    OUTPUT_INSTRUCTION_EXPLANATION,
     UNSUPPORTED_PROMPT_TYPE_ERROR,
 )
 from .enums import BlockType
@@ -176,15 +177,14 @@ class MarkdownFlow:
 
     # Core unified interface
 
-    def process(
+    async def process(
         self,
         block_index: int,
         mode: ProcessMode = ProcessMode.COMPLETE,
         context: list[dict[str, str]] | None = None,
         variables: dict[str, str | list[str]] | None = None,
         user_input: dict[str, list[str]] | None = None,
-        dynamic_interaction_format: str | None = None,
-    ) -> LLMResult | Generator[LLMResult, None, None]:
+    ) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """
         Unified block processing interface.
 
@@ -194,10 +194,9 @@ class MarkdownFlow:
             context: Context message list
             variables: Variable mappings
             user_input: User input (for interaction blocks)
-            dynamic_interaction_format: Dynamic interaction format for validation
 
         Returns:
-            LLMResult or Generator[LLMResult, None, None]
+            LLMResult or AsyncGenerator[LLMResult, None]
         """
         # Process document_prompt variable replacement
         if self._document_prompt:
@@ -206,70 +205,56 @@ class MarkdownFlow:
         block = self.get_block(block_index)
 
         if block.block_type == BlockType.CONTENT:
-            # Check if this is dynamic interaction validation
-            if dynamic_interaction_format and user_input:
-                return self._process_dynamic_interaction_validation(block_index, dynamic_interaction_format, user_input, mode, context, variables)
-            # Normal content processing (possibly with dynamic conversion)
-            return self._process_content(block_index, mode, context, variables)
+            return await self._process_content(block_index, mode, context, variables)
 
         if block.block_type == BlockType.INTERACTION:
             if user_input is None:
                 # Render interaction content
-                return self._process_interaction_render(block_index, mode, variables)
+                return await self._process_interaction_render(block_index, mode, variables)
             # Process user input
-            return self._process_interaction_input(block_index, user_input, mode, context, variables)
+            return await self._process_interaction_input(block_index, user_input, mode, context, variables)
 
         if block.block_type == BlockType.PRESERVED_CONTENT:
             # Preserved content output as-is, no LLM call
-            return self._process_preserved_content(block_index, variables)
+            return await self._process_preserved_content(block_index, variables)
 
         # Handle other types as content
-        return self._process_content(block_index, mode, context, variables)
+        return await self._process_content(block_index, mode, context, variables)
 
     # Internal processing methods
 
-    def _process_content(
+    async def _process_content(
         self,
         block_index: int,
         mode: ProcessMode,
         context: list[dict[str, str]] | None,
         variables: dict[str, str | list[str]] | None,
-    ) -> LLMResult | Generator[LLMResult, None, None]:
+    ) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """Process content block."""
+        # Build messages
+        messages = self._build_content_messages(block_index, variables)
 
-        # For PROMPT_ONLY mode, use standard content processing
         if mode == ProcessMode.PROMPT_ONLY:
-            messages = self._build_content_messages(block_index, variables)
             return LLMResult(prompt=messages[-1]["content"], metadata={"messages": messages})
 
-        # For COMPLETE and STREAM modes with LLM provider, use dynamic interaction check
-        # LLM will decide whether content needs to be converted to interaction block
-        if self._llm_provider:
-            block = self.get_block(block_index)
-            if block.block_type == BlockType.CONTENT:
-                return self._process_with_dynamic_check(block_index, mode, context, variables)
-
-        # Fallback: Build messages using standard content processing
-        messages = self._build_content_messages(block_index, variables)
-
         if mode == ProcessMode.COMPLETE:
             if not self._llm_provider:
                 raise ValueError(LLM_PROVIDER_REQUIRED_ERROR)
 
-            result = self._llm_provider.complete(messages)
-            return LLMResult(content=result.content, prompt=messages[-1]["content"], metadata=result.metadata)
+            content = await self._llm_provider.complete(messages)
+            return LLMResult(content=content, prompt=messages[-1]["content"])
 
         if mode == ProcessMode.STREAM:
             if not self._llm_provider:
                 raise ValueError(LLM_PROVIDER_REQUIRED_ERROR)
 
-            def stream_generator():
-                for chunk in self._llm_provider.stream(messages):
+            async def stream_generator():
+                async for chunk in self._llm_provider.stream(messages):  # type: ignore[attr-defined]
                     yield LLMResult(content=chunk, prompt=messages[-1]["content"])
 
             return stream_generator()
 
-    def _process_preserved_content(self, block_index: int, variables: dict[str, str | list[str]] | None) -> LLMResult:
+    async def _process_preserved_content(self, block_index: int, variables: dict[str, str | list[str]] | None) -> LLMResult:
         """Process preserved content block, output as-is without LLM call."""
         block = self.get_block(block_index)
 
@@ -281,7 +266,7 @@ class MarkdownFlow:
 
         return LLMResult(content=content)
 
-    def _process_interaction_render(self, block_index: int, mode: ProcessMode, variables: dict[str, str | list[str]] | None = None) -> LLMResult | Generator[LLMResult, None, None]:
+    async def _process_interaction_render(self, block_index: int, mode: ProcessMode, variables: dict[str, str | list[str]] | None = None) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """Process interaction content rendering."""
         block = self.get_block(block_index)
 
@@ -314,8 +299,7 @@ class MarkdownFlow:
             if not self._llm_provider:
                 return LLMResult(content=processed_block.content)  # Fallback processing
 
-            result = self._llm_provider.complete(messages)
-            rendered_question = result.content
+            rendered_question = await self._llm_provider.complete(messages)
             rendered_content = self._reconstruct_interaction_content(processed_block.content, rendered_question)
 
             return LLMResult(
@@ -332,7 +316,7 @@ class MarkdownFlow:
                 # For interaction blocks, return reconstructed content (one-time output)
                 rendered_content = self._reconstruct_interaction_content(processed_block.content, question_text or "")
 
-                def stream_generator():
+                async def stream_generator():
                     yield LLMResult(
                         content=rendered_content,
                         prompt=messages[-1]["content"],
@@ -341,9 +325,9 @@ class MarkdownFlow:
                 return stream_generator()
 
             # With LLM provider, collect full response then return once
-            def stream_generator():
+            async def stream_generator():
                 full_response = ""
-                for chunk in self._llm_provider.stream(messages):
+                async for chunk in self._llm_provider.stream(messages):  # type: ignore[attr-defined]
                     full_response += chunk
 
                 # Reconstruct final interaction content
@@ -357,23 +341,22 @@ class MarkdownFlow:
 
             return stream_generator()
 
-    def _process_interaction_input(
+    async def _process_interaction_input(
         self,
         block_index: int,
         user_input: dict[str, list[str]],
         mode: ProcessMode,
         context: list[dict[str, str]] | None,
         variables: dict[str, str | list[str]] | None = None,
-    ) -> LLMResult | Generator[LLMResult, None, None]:
+    ) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """Process interaction user input."""
-        _ = context  # Mark as intentionally unused
         block = self.get_block(block_index)
         target_variable = block.variables[0] if block.variables else "user_input"
 
         # Basic validation
         if not user_input or not any(values for values in user_input.values()):
             error_msg = INPUT_EMPTY_ERROR
-            return self._render_error(error_msg, mode)
+            return await self._render_error(error_msg, mode)
 
         # Get the target variable value from user_input
         target_values = user_input.get(target_variable, [])
@@ -387,7 +370,7 @@ class MarkdownFlow:
 
         if "error" in parse_result:
             error_msg = INTERACTION_PARSE_ERROR.format(error=parse_result["error"])
-            return self._render_error(error_msg, mode)
+            return await self._render_error(error_msg, mode)
 
         interaction_type = parse_result.get("type")
 
@@ -399,7 +382,7 @@ class MarkdownFlow:
             InteractionType.BUTTONS_MULTI_WITH_TEXT,
         ]:
             # All button types: validate user input against available buttons
-            return self._process_button_validation(
+            return await self._process_button_validation(
                 parse_result,
                 target_values,
                 target_variable,
@@ -432,16 +415,16 @@ class MarkdownFlow:
                 },
             )
         error_msg = f"No input provided for variable '{target_variable}'"
-        return self._render_error(error_msg, mode)
+        return await self._render_error(error_msg, mode)
 
-    def _process_button_validation(
+    async def _process_button_validation(
         self,
         parse_result: dict[str, Any],
         target_values: list[str],
         target_variable: str,
         mode: ProcessMode,
         interaction_type: InteractionType,
-    ) -> LLMResult | Generator[LLMResult, None, None]:
+    ) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """
         Simplified button validation with new input format.
 
@@ -476,9 +459,9 @@ class MarkdownFlow:
             # Pure button mode requires input
             button_displays = [btn["display"] for btn in buttons]
             error_msg = f"Please select from: {', '.join(button_displays)}"
-            return self._render_error(error_msg, mode)
+            return await self._render_error(error_msg, mode)
 
-        # First, check if user input matches available buttons
+        # Validate input values against available buttons
         valid_values = []
         invalid_values = []
 
@@ -491,30 +474,19 @@ class MarkdownFlow:
                     break
 
             if not matched:
-                invalid_values.append(value)
-
-        # If there are invalid values and this interaction allows text input, use LLM validation
-        if invalid_values and allow_text_input:
-            # Use LLM validation for text input interactions
-            button_displays = [btn["display"] for btn in buttons]
-            question = parse_result.get("question", "")
-
-            return self._process_llm_validation_with_options(
-                block_index=0,  # Not used in the method
-                user_input={target_variable: target_values},
-                target_variable=target_variable,
-                options=button_displays,
-                question=question,
-                mode=mode,
-            )
-
-        # Check for validation errors in pure button mode or when text input not allowed
-        if invalid_values:
+                if allow_text_input:
+                    # Allow custom text in buttons+text mode
+                    valid_values.append(value)
+                else:
+                    invalid_values.append(value)
+
+        # Check for validation errors
+        if invalid_values and not allow_text_input:
             button_displays = [btn["display"] for btn in buttons]
             error_msg = f"Invalid options: {', '.join(invalid_values)}. Please select from: {', '.join(button_displays)}"
-            return self._render_error(error_msg, mode)
+            return await self._render_error(error_msg, mode)
 
-        # Success: return validated button values
+        # Success: return validated values
         return LLMResult(
             content="",
             variables={target_variable: valid_values},
@@ -524,17 +496,16 @@ class MarkdownFlow:
                 "valid_values": valid_values,
                 "invalid_values": invalid_values,
                 "total_input_count": len(target_values),
-                "llm_validated": False,
             },
         )
 
-    def _process_llm_validation(
+    async def _process_llm_validation(
         self,
         block_index: int,
         user_input: dict[str, list[str]],
         target_variable: str,
         mode: ProcessMode,
-    ) -> LLMResult | Generator[LLMResult, None, None]:
+    ) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """Process LLM validation."""
         # Build validation messages
         messages = self._build_validation_messages(block_index, user_input, target_variable)
@@ -553,8 +524,7 @@ class MarkdownFlow:
                 # Fallback processing, return variables directly
                 return LLMResult(content="", variables=user_input)  # type: ignore[arg-type]
 
-            result = self._llm_provider.complete(messages)
-            llm_response = result.content
+            llm_response = await self._llm_provider.complete(messages)
 
             # Parse validation response and convert to LLMResult
             # Use joined target values for fallback; avoids JSON string injection
@@ -566,9 +536,9 @@ class MarkdownFlow:
             if not self._llm_provider:
                 return LLMResult(content="", variables=user_input)  # type: ignore[arg-type]
 
-            def stream_generator():
+            async def stream_generator():
                 full_response = ""
-                for chunk in self._llm_provider.stream(messages):
+                async for chunk in self._llm_provider.stream(messages):  # type: ignore[attr-defined]
                     full_response += chunk
 
                 # Parse complete response and convert to LLMResult
@@ -582,7 +552,7 @@ class MarkdownFlow:
 
             return stream_generator()
 
-    def _process_llm_validation_with_options(
+    async def _process_llm_validation_with_options(
         self,
         block_index: int,
         user_input: dict[str, list[str]],
@@ -590,9 +560,8 @@ class MarkdownFlow:
         options: list[str],
         question: str,
         mode: ProcessMode,
-    ) -> LLMResult | Generator[LLMResult, None, None]:
+    ) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """Process LLM validation with button options (third case)."""
-        _ = block_index  # Mark as intentionally unused
         # Build special validation messages containing button option information
         messages = self._build_validation_messages_with_options(user_input, target_variable, options, question)
 
@@ -612,8 +581,7 @@ class MarkdownFlow:
                 # Fallback processing, return variables directly
                 return LLMResult(content="", variables=user_input)  # type: ignore[arg-type]
 
-            result = self._llm_provider.complete(messages)
-            llm_response = result.content
+            llm_response = await self._llm_provider.complete(messages)
 
             # Parse validation response and convert to LLMResult
             # Use joined target values for fallback; avoids JSON string injection
@@ -625,9 +593,9 @@ class MarkdownFlow:
             if not self._llm_provider:
                 return LLMResult(content="", variables=user_input)  # type: ignore[arg-type]
 
-            def stream_generator():
+            async def stream_generator():
                 full_response = ""
-                for chunk in self._llm_provider.stream(messages):
+                async for chunk in self._llm_provider.stream(messages):  # type: ignore[attr-defined]
                     full_response += chunk
                     # For validation scenario, don't output chunks in real-time, only final result
 
@@ -644,7 +612,7 @@ class MarkdownFlow:
 
             return stream_generator()
 
-    def _render_error(self, error_message: str, mode: ProcessMode) -> LLMResult | Generator[LLMResult, None, None]:
+    async def _render_error(self, error_message: str, mode: ProcessMode) -> LLMResult | AsyncGenerator[LLMResult, None]:
         """Render user-friendly error message."""
         messages = self._build_error_render_messages(error_message)
 
@@ -658,16 +626,15 @@ class MarkdownFlow:
             if not self._llm_provider:
                 return LLMResult(content=error_message)  # Fallback processing
 
-            result = self._llm_provider.complete(messages)
-            friendly_error = result.content
+            friendly_error = await self._llm_provider.complete(messages)
             return LLMResult(content=friendly_error, prompt=messages[-1]["content"])
 
         if mode == ProcessMode.STREAM:
             if not self._llm_provider:
                 return LLMResult(content=error_message)
 
-            def stream_generator():
-                for chunk in self._llm_provider.stream(messages):
+            async def stream_generator():
+                async for chunk in self._llm_provider.stream(messages):  # type: ignore[attr-defined]
                     yield LLMResult(content=chunk, prompt=messages[-1]["content"])
 
             return stream_generator()
@@ -683,8 +650,9 @@ class MarkdownFlow:
         block = self.get_block(block_index)
         block_content = block.content
 
-        # Process output instructions
-        block_content = process_output_instructions(block_content)
+        # Process output instructions and detect if preserved content exists
+        # Returns: (processed_content, has_preserved_content)
+        block_content, has_preserved_content = process_output_instructions(block_content)
 
         # Replace variables
         block_content = replace_variables_in_text(block_content, variables or {})
@@ -692,9 +660,16 @@ class MarkdownFlow:
         # Build message array
         messages = []
 
-        # Add document prompt
+        # Conditionally add system prompts
         if self._document_prompt:
-            messages.append({"role": "system", "content": self._document_prompt})
+            system_msg = self._document_prompt
+            # Only add output instruction explanation when preserved content detected
+            if has_preserved_content:
+                system_msg += "\n\n" + OUTPUT_INSTRUCTION_EXPLANATION.strip()
+            messages.append({"role": "system", "content": system_msg})
+        elif has_preserved_content:
+            # No document prompt but has preserved content, add explanation alone
+            messages.append({"role": "system", "content": OUTPUT_INSTRUCTION_EXPLANATION.strip()})
 
         # For most content blocks, historical conversation context is not needed
         # because each document block is an independent instruction
@@ -827,411 +802,5 @@ Original Error: {error_message}
         if match:
             prefix = match.group(1)
             suffix = match.group(2)
-            # Extract only the closing bracket from suffix, remove original question
-            # suffix format is "original_question]", we only want "]"
-            if suffix.endswith("]"):
-                clean_suffix = "]"
-            else:
-                clean_suffix = suffix
-
-            return f"{prefix}{cleaned_question}{clean_suffix}"
+            return f"{prefix}{cleaned_question}{suffix}"
         return original_content  # type: ignore[unreachable]
-
-    # Dynamic Interaction Methods
-
-    def _process_with_dynamic_check(
-        self,
-        block_index: int,
-        mode: ProcessMode,
-        context: list[dict[str, str]] | None,
-        variables: dict[str, str | list[str]] | None,
-    ) -> LLMResult | Generator[LLMResult, None, None]:
-        """Process content with dynamic interaction detection and conversion."""
-
-        block = self.get_block(block_index)
-        messages = self._build_dynamic_check_messages(block, context, variables)
-
-        # Define Function Calling tools with structured approach
-        tools = [
-            {
-                "type": "function",
-                "function": {
-                    "name": "create_interaction_block",
-                    "description": "Convert content to interaction block with structured data when it needs to collect user input",
-                    "parameters": {
-                        "type": "object",
-                        "properties": {
-                            "needs_interaction": {"type": "boolean", "description": "Whether this content needs to be converted to interaction block"},
-                            "variable_name": {"type": "string", "description": "Name of the variable to collect (without {{}} brackets)"},
-                            "interaction_type": {
-                                "type": "string",
-                                "enum": ["single_select", "multi_select", "text_input", "mixed"],
-                                "description": "Type of interaction: single_select (|), multi_select (||), text_input (...), mixed (options + text)",
-                            },
-                            "options": {"type": "array", "items": {"type": "string"}, "description": "List of selectable options (3-4 specific options based on context)"},
-                            "allow_text_input": {"type": "boolean", "description": "Whether to include a text input option for 'Other' cases"},
-                            "text_input_prompt": {"type": "string", "description": "Prompt text for the text input option (e.g., '其他请输入', 'Other, please specify')"},
-                        },
-                        "required": ["needs_interaction"],
-                    },
-                },
-            }
-        ]
-
-        if not self._llm_provider:
-            raise ValueError(LLM_PROVIDER_REQUIRED_ERROR)
-
-        # Call LLM with tools
-        result = self._llm_provider.complete(messages, tools)
-
-        # If interaction was generated through Function Calling, construct the MarkdownFlow format
-        if result.transformed_to_interaction and result.metadata and "tool_args" in result.metadata:
-            tool_args = result.metadata["tool_args"]
-            if tool_args.get("needs_interaction"):
-                # Construct MarkdownFlow format from structured data
-                interaction_content = self._build_interaction_format(tool_args)
-                result.content = interaction_content
-
-        # If transformed to interaction, return as is
-        if result.transformed_to_interaction:
-            return result
-
-        # If not transformed, continue with normal processing using standard content messages
-        normal_messages = self._build_content_messages(block_index, variables)
-
-        if mode == ProcessMode.STREAM:
-
-            def stream_wrapper():
-                stream_generator = self._llm_provider.stream(normal_messages)
-                for chunk in stream_generator:
-                    yield LLMResult(content=chunk)
-
-            return stream_wrapper()
-
-        # Complete mode - use normal content processing
-        normal_result = self._llm_provider.complete(normal_messages)
-        return LLMResult(content=normal_result.content, prompt=normal_messages[-1]["content"], metadata=normal_result.metadata)
-
-    def _build_dynamic_check_messages(
-        self,
-        block: "Block",
-        context: list[dict[str, str]] | None,
-        variables: dict[str, str | list[str]] | None,
-    ) -> list[dict[str, str]]:
-        """Build messages for dynamic interaction detection."""
-
-        import json
-
-        # System prompt for detection
-        system_prompt = """You are an intelligent document processing assistant specializing in creating interactive forms.
-
-Task: Analyze the given content block and determine if it needs to be converted to an interaction block to collect user information.
-
-**ABSOLUTE RULE**: Convert ONLY when ALL THREE mandatory elements are explicitly present:
-1. Storage action word + target connector + variable
-2. No exceptions, no implications, no assumptions
-
-**MANDATORY TRIPLE PATTERN (ALL REQUIRED):**
-
-**Element 1: Storage Action Words**
-- Chinese: "记录", "保存", "存储", "收集", "采集"
-- English: "save", "store", "record", "collect", "gather"
-
-**Element 2: Target Connection Words**
-- Chinese: "到", "为", "在", "至"
-- English: "to", "as", "in", "into"
-
-**Element 3: Target Variable**
-- Must contain {{variable_name}} syntax for NEW data storage
-- Variable must be for collecting NEW information, not using existing data
-
-**VALID CONVERSION FORMULA:**
-[Storage Word] + [Connector] + {{new_variable}}
-
-Examples of VALID patterns:
-- "...记录到{{姓名}}"
-- "...保存为{{偏好}}"
-- "...存储在{{选择}}"
-- "...save to {{preference}}"
-- "...collect as {{user_input}}"
-
-**STRICT EXCLUSION RULES:**
-
-❌ NEVER convert if missing ANY element:
-- No storage action word = NO conversion
-- No target connector = NO conversion
-- No {{variable}} = NO conversion
-- Using existing {{variable}} instead of collecting new = NO conversion
-
-❌ NEVER convert casual conversation:
-- Simple questions without storage intent
-- Introduction requests without persistence
-- General inquiries without data collection
-- Educational or exploratory content
-
-❌ NEVER infer or assume storage intent:
-- Don't assume "询问姓名" means "保存姓名"
-- Don't assume "了解偏好" means "记录偏好"
-- Don't assume data collection without explicit storage words
-
-**PATTERN ANALYSIS METHOD:**
-1. **Exact Pattern Match**: Search for [Storage Word] + [Connector] + {{variable}}
-2. **No Pattern = No Conversion**: If exact pattern not found, return needs_interaction: false
-3. **Zero Tolerance**: No partial matches, no similar meanings, no interpretations
-
-**ULTRA-CONSERVATIVE APPROACH:**
-- If there's ANY doubt about storage intent = DON'T convert
-- If storage pattern is not 100% explicit = DON'T convert
-- If you need to "interpret" or "infer" storage intent = DON'T convert
-- Prefer false negatives over false positives
-
-When exact pattern is found, generate structured interaction data. Otherwise, always return needs_interaction: false."""
-
-        # User message with content and context
-        # Build user prompt with document context
-        user_prompt_parts = []
-
-        # Add document-level prompt context if exists
-        if self._document_prompt:
-            user_prompt_parts.append(f"""Document-level instructions:
-{self._document_prompt}
-
-(Note: The above are the user's document-level instructions that provide context and requirements for processing.)
-""")
-
-        # Prepare content analysis with both original and resolved versions
-        original_content = block.content
-
-        # Create resolved content with variable substitution for better context
-        resolved_content = original_content
-        if variables:
-            from .utils import replace_variables_in_text
-
-            resolved_content = replace_variables_in_text(original_content, variables)
-
-        content_analysis = f"""Current content block to analyze:
-
-**Original content (shows variable structure):**
-{original_content}
-
-**Resolved content (with current variable values):**
-{resolved_content}
-
-**Existing variable values:**
-{json.dumps(variables, ensure_ascii=False) if variables else "None"}"""
-
-        # Add different analysis based on whether content has variables
-        if "{{" in original_content and "}}" in original_content:
-            from .utils import extract_variables_from_text
-
-            content_variables = set(extract_variables_from_text(original_content))
-
-            # Find new variables (not yet collected)
-            new_variables = content_variables - (set(variables.keys()) if variables else set())
-            existing_used_variables = content_variables & (set(variables.keys()) if variables else set())
-
-            content_analysis += f"""
-
-**Variable analysis:**
-- Variables used from previous steps: {list(existing_used_variables) if existing_used_variables else "None"}
-- New variables to collect: {list(new_variables) if new_variables else "None"}
-
-**Context guidance:**
-- Use the resolved content to understand the actual context and requirements
-- Generate options based on the real variable values shown in the resolved content
-- Collect user input for the new variables identified above"""
-
-        user_prompt_parts.append(content_analysis)
-
-        # Add analysis requirements and structured output guide
-        user_prompt_parts.append("""## Analysis Task:
-1. Determine if this content needs to be converted to an interaction block
-2. If conversion is needed, provide structured interaction data
-
-## Context-based Analysis:
-- Use the "Resolved content" to understand actual context (e.g., if it shows "川菜", generate Sichuan dish options)
-- Extract the "New variables to collect" identified in the variable analysis above
-- Generate 3-4 specific options based on the resolved context and document-level instructions
-- Follow ALL document-level instruction requirements (language, domain, terminology)
-
-## Selection Type Decision Logic:
-Ask: "Can a user realistically want/choose multiple of these options simultaneously?"
-
-**Use MULTI_SELECT when:**
-- Food dishes (can order multiple: 宫保鸡丁, 麻婆豆腐)
-- Programming skills (can know multiple: Python, JavaScript)
-- Interests/hobbies (can have multiple: 读书, 运动, 旅游)
-- Product features (can want multiple: 定制颜色, 个性化logo)
-- Exercise types (can do multiple: 跑步, 游泳, 瑜伽)
-
-**Use SINGLE_SELECT when:**
-- Job positions (usually apply for one: 软件工程师 OR 产品经理)
-- Experience levels (have one current level: Beginner OR Advanced)
-- Budget ranges (have one range: 5-10万 OR 10-20万)
-- Education levels (have one highest: Bachelor's OR Master's)
-
-## Output Instructions:
-If this content needs interaction, use the create_interaction_block function with:
-- `needs_interaction`: true/false
-- `variable_name`: the variable to collect (from "New variables" above)
-- `interaction_type`: "single_select", "multi_select", "text_input", or "mixed"
-- `options`: array of 3-4 specific options based on context
-- `allow_text_input`: true if you want to include "other" option
-- `text_input_prompt`: text for the "other" option (in appropriate language)
-
-Analyze the content and provide the structured interaction data.""")
-
-        user_prompt = "\n\n".join(user_prompt_parts)
-
-        messages = [{"role": "system", "content": system_prompt}]
-
-        # Add context if provided
-        if context:
-            messages.extend(context)
-
-        messages.append({"role": "user", "content": user_prompt})
-
-        return messages
-
-    def _build_interaction_format(self, tool_args: dict) -> str:
-        """Build MarkdownFlow interaction format from structured Function Calling data."""
-        variable_name = tool_args.get("variable_name", "")
-        interaction_type = tool_args.get("interaction_type", "single_select")
-        options = tool_args.get("options", [])
-        allow_text_input = tool_args.get("allow_text_input", False)
-        text_input_prompt = tool_args.get("text_input_prompt", "...请输入")
-
-        if not variable_name:
-            return ""
-
-        # For text_input type, options can be empty
-        if interaction_type != "text_input" and not options:
-            return ""
-
-        # Choose separator based on interaction type
-        if interaction_type in ["multi_select", "mixed"]:
-            separator = "||"
-        else:
-            separator = "|"
-
-        # Build options string
-        if interaction_type == "text_input":
-            # Text input only
-            options_str = f"...{text_input_prompt}"
-        else:
-            # Options with potential text input
-            options_str = separator.join(options)
-
-            if allow_text_input and text_input_prompt:
-                # Ensure text input has ... prefix
-                text_option = text_input_prompt if text_input_prompt.startswith("...") else f"...{text_input_prompt}"
-                options_str += f"{separator}{text_option}"
-
-        return f"?[%{{{{{variable_name}}}}} {options_str}]"
-
-    def _process_dynamic_interaction_validation(
-        self,
-        block_index: int,
-        interaction_format: str,
-        user_input: dict[str, list[str]],
-        mode: ProcessMode,
-        context: list[dict[str, str]] | None,
-        variables: dict[str, str | list[str]] | None,
-    ) -> LLMResult:
-        """Validate user input for dynamically generated interaction blocks using same logic as normal interactions."""
-        _ = block_index  # Mark as intentionally unused
-        _ = context  # Mark as intentionally unused
-
-        from .utils import InteractionParser
-
-        # Parse the interaction format using the same parser as normal interactions
-        parser = InteractionParser()
-        parse_result = parser.parse(interaction_format)
-
-        if "error" in parse_result:
-            error_msg = f"Invalid interaction format: {parse_result['error']}"
-            return self._render_error(error_msg, mode)
-
-        # Extract variable name and interaction type
-        variable_name = parse_result.get("variable")
-        interaction_type = parse_result.get("type")
-
-        if not variable_name:
-            error_msg = f"No variable found in interaction format: {interaction_format}"
-            return self._render_error(error_msg, mode)
-
-        # Get user input for the target variable
-        target_values = user_input.get(variable_name, [])
-
-        # Basic validation - check if input is provided when required
-        if not target_values:
-            # Check if this is a text input or allows empty input
-            allow_text_input = interaction_type in [
-                InteractionType.BUTTONS_WITH_TEXT,
-                InteractionType.BUTTONS_MULTI_WITH_TEXT,
-            ]
-
-            if allow_text_input:
-                # Allow empty input for buttons+text mode - merge with existing variables
-                merged_variables = dict(variables or {})
-                merged_variables[variable_name] = []
-                return LLMResult(
-                    content="",
-                    variables=merged_variables,
-                    metadata={
-                        "interaction_type": "dynamic_interaction",
-                        "empty_input": True,
-                    },
-                )
-            error_msg = f"No input provided for variable '{variable_name}'"
-            return self._render_error(error_msg, mode)
-
-        # Use the same validation logic as normal interactions
-        if interaction_type in [
-            InteractionType.BUTTONS_ONLY,
-            InteractionType.BUTTONS_WITH_TEXT,
-            InteractionType.BUTTONS_MULTI_SELECT,
-            InteractionType.BUTTONS_MULTI_WITH_TEXT,
-        ]:
-            # Button validation - reuse the existing button validation logic
-            button_result = self._process_button_validation(
-                parse_result,
-                target_values,
-                variable_name,
-                mode,
-                interaction_type,
-            )
-
-            # Merge with existing variables for dynamic interactions
-            if hasattr(button_result, "variables") and button_result.variables is not None and variables:
-                merged_variables = dict(variables)
-                merged_variables.update(button_result.variables)
-                return LLMResult(
-                    content=button_result.content,
-                    variables=merged_variables,
-                    metadata=button_result.metadata,
-                )
-            return button_result
-
-        if interaction_type == InteractionType.NON_ASSIGNMENT_BUTTON:
-            # Non-assignment buttons: don't set variables, keep existing ones
-            return LLMResult(
-                content="",
-                variables=dict(variables or {}),
-                metadata={
-                    "interaction_type": "non_assignment_button",
-                    "user_input": user_input,
-                },
-            )
-        # Text-only input type - merge with existing variables
-        merged_variables = dict(variables or {})
-        merged_variables[variable_name] = target_values
-        return LLMResult(
-            content="",
-            variables=merged_variables,
-            metadata={
-                "interaction_type": "text_only",
-                "target_variable": variable_name,
-                "values": target_values,
-            },
-        )