Added provider-specific prompt infrastructure; thinking and progress indicators to chat ui

app.py

@@ -691,7 +691,8 @@ async def chat_with_tool_visibility(
     openai_key,
     anthropic_key,
     agent_mode,
-    request: gr.Request
+    request: gr.Request,
+    progress=gr.Progress()
 ):
     """
     Dual-output streaming: chat response + tool execution log
@@ -729,6 +730,8 @@ async def chat_with_tool_visibility(
     # -------------------------------------------------------------------------
     # 2. GET OR CREATE AGENT
     # -------------------------------------------------------------------------
+    progress(0.1, desc="🔧 Initializing agent...")
+
     try:
         session_id = request.session_hash
 
@@ -750,6 +753,8 @@ async def chat_with_tool_visibility(
         yield f"**Agent Creation Failed**\n\n{str(e)}", "*Agent creation failed*"
         return
 
+    progress(0.3, desc="🤖 Agent ready...")
+
     config = {"configurable": {"thread_id": session_id}}
 
     # -------------------------------------------------------------------------
@@ -817,6 +822,12 @@ async def chat_with_tool_visibility(
     # -------------------------------------------------------------------------
     # 4. STREAM AGENT RESPONSE WITH TOOL VISIBILITY
     # -------------------------------------------------------------------------
+    # Initial "thinking" indicator
+    progress(0.5, desc="💭 Thinking...")
+    chat_response = "💭 _Thinking..._"
+    tool_log += "🔵 Agent started processing...\n"
+    yield chat_response, tool_log
+
     print(f"[DEBUG AGENT INPUT] Sending to agent: {user_text}")  # DEBUG
     async for event in agent.astream_events(
         {"messages": [{"role": "user", "content": user_text}]},
@@ -831,6 +842,9 @@ async def chat_with_tool_visibility(
             tool_name = event["name"]
             tool_input = event.get("data", {}).get("input", {})
 
+            # Update progress
+            progress(0.6 + (tool_count * 0.05), desc=f"🔍 Using {tool_name}...")
+
             # Add to tool log
             tool_log += f"\n🟢 Tool #{tool_count}: {tool_name}\n"
             tool_log += f"Status: Running...\n"
@@ -845,6 +859,11 @@ async def chat_with_tool_visibility(
         elif kind == "on_chat_model_stream":
             content = event["data"]["chunk"].content
             if content:
+                # Clear "Thinking..." on first real content
+                if chat_response == "💭 _Thinking..._":
+                    chat_response = ""
+                    progress(0.7, desc="📝 Generating response...")
+
                 # Handle both string (OpenAI) and list (Anthropic) content formats
                 if isinstance(content, list):
                     # Anthropic returns list of content blocks - extract text
@@ -862,6 +881,9 @@ async def chat_with_tool_visibility(
         elif kind == "on_tool_end":
             tool_output = event.get("data", {}).get("output", "")
 
+            # Update progress
+            progress(0.8, desc="📊 Processing results...")
+
             # Format output for tool log (truncate if needed)
             output_str = str(tool_output)
             if len(output_str) > 1000:
@@ -879,9 +901,11 @@ async def chat_with_tool_visibility(
 
             yield chat_response, tool_log
 
-    # Final yield 
+    # Final yield
     ## NEW: Updated with LlamaIndex OutputPraser
     # yield chat_response, tool_log
+    progress(0.9, desc="✨ Finalizing response...")
+
     try:
         from langgraph_agent.structured_output import parse_agent_response
         formatted_response = await parse_agent_response(
@@ -890,13 +914,16 @@ async def chat_with_tool_visibility(
             api_key=api_key,
             model=model
         )
+        progress(1.0, desc="✅ Complete")
         yield formatted_response, tool_log
     except ImportError:
         # Fallback if LlamaIndex not installed
+        progress(1.0, desc="✅ Complete")
         yield chat_response, tool_log
     except Exception as e:
         # Fallback if parsing fails
         print(f"[STRUCTURED OUTPUT ERROR]: {e}")
+        progress(1.0, desc="✅ Complete")
         yield chat_response, tool_log
 
 
@@ -947,7 +974,7 @@ async def check_modal_server_health():
             return f"❌ Offline"
 
 # Wrapper to convert to Gradio 6 message format
-async def chat_wrapper(message, history, provider, hf_key, openai_key, anthropic_key, agent_mode, tool_log_state, request: gr.Request):
+async def chat_wrapper(message, history, provider, hf_key, openai_key, anthropic_key, agent_mode, tool_log_state, request: gr.Request, progress=gr.Progress()):
     """
     Wrapper to convert chat outputs to Gradio 6 message format.
 
@@ -966,7 +993,7 @@ async def chat_wrapper(message, history, provider, hf_key, openai_key, anthropic
     history = history + [{"role": "user", "content": user_message_text}]
 
     # Stream response
-    async for chat_text, tool_log_text in chat_with_tool_visibility(message, history, provider, hf_key, openai_key, anthropic_key, agent_mode, request):
+    async for chat_text, tool_log_text in chat_with_tool_visibility(message, history, provider, hf_key, openai_key, anthropic_key, agent_mode, request, progress):
         # Update history with assistant response
         updated_history = history + [{"role": "assistant", "content": chat_text}]
         yield updated_history, tool_log_text

docs/dev/prompts_config-README.md

@@ -0,0 +1,98 @@
+# How Developers Work With This System
+
+## 1. Adding a New Provider (e.g., "Claude" prompts)
+
+### Step 1: Create the prompt in prompts.py:
+
+```python
+# Add new prompt variant
+AUDIO_FINDER_PROMPT_CLAUDE = """You are BirdScope Audio Finder optimized for Claude..."""
+
+# Update PROMPTS dict
+PROMPTS = {
+    "audio_finder": {
+        "default": AUDIO_FINDER_PROMPT,
+        "huggingface": AUDIO_FINDER_PROMPT_HF,
+        "claude": AUDIO_FINDER_PROMPT_CLAUDE,  # NEW
+    },
+}
+```
+
+That's it! The system automatically picks it up when `provider="claude"` is passed.
+
+## 2. Adding a New Prompt Type (e.g., "data_analyst")
+
+### Step 1: Create prompts:
+
+```python
+DATA_ANALYST_PROMPT = """Default data analyst prompt..."""
+DATA_ANALYST_PROMPT_HF = """HF-optimized data analyst prompt..."""
+
+PROMPTS = {
+    # ... existing prompts ...
+    "data_analyst": {
+        "default": DATA_ANALYST_PROMPT,
+        "huggingface": DATA_ANALYST_PROMPT_HF,
+    }
+}
+```
+
+### Step 2: Use in subagent_config.py:
+
+```python
+"data_analyst": {
+    "name": "Data Analyst",
+    "tools": [...],
+    "prompt": get_prompt("data_analyst", provider) or DATA_ANALYST_PROMPT,
+}
+```
+
+## 3. Testing Different Prompts
+
+### Option A: Through UI (current method)
+- Run `python app.py`
+- Select provider dropdown → "HuggingFace"
+- System automatically uses HF prompts
+
+### Option B: Programmatically (for testing)
+
+```python
+from langgraph_agent import prompts
+
+# Test which prompt is selected
+prompt = prompts.get_prompt("audio_finder", "huggingface")
+print(f"Length: {len(prompt)}")
+print(prompt[:100])
+```
+
+## 4. Fallback Behavior
+
+The system is designed with safe fallbacks:
+
+```python
+# If HuggingFace variant doesn't exist, falls back to default
+prompt = get_prompt("species_explorer", "huggingface")
+# Returns: SPECIES_EXPLORER_PROMPT_HF if exists, else None
+
+# In subagent_config.py, the "or" ensures a default
+prompt = get_prompt("species_explorer", provider) or """Inline default..."""
+```
+
+## Developer Workflow Summary
+
+### To modify prompts:
+1. Edit `langgraph_agent/prompts.py`
+2. Add/modify prompt strings
+3. Update `PROMPTS` dictionary
+4. Restart app - changes take effect immediately
+
+### To add new provider support:
+1. Add provider key to `PROMPTS` dict
+2. No other changes needed - fallback handles missing variants
+
+### To debug which prompt is used:
+
+```python
+from langgraph_agent.prompts import get_prompt
+print(get_prompt("audio_finder", "huggingface")[:200])
+```

langgraph_agent/agents.py

@@ -73,7 +73,7 @@ class AgentFactory:
         if mode_config["use_router"]:
             # Multi-agent mode: create router with specialists
             print(f"[AGENT]: Creating supervisor with subagents: {mode_config['subagents']}")
-            workflow = await create_supervisor_workflow(tools, llm)
+            workflow = await create_supervisor_workflow(tools, llm, provider=provider)
             return workflow
         else:
             # Single agent mode: create one subagent directly
@@ -86,14 +86,16 @@ class AgentFactory:
 
             # create_agent auto-compiles, so pass checkpointer and name directly
             # Filter tools based on subagent configuration
-            subagent_tools = SubAgentConfig.get_subagent_definitions()["generalist"]["tools"]
+            # Pass provider to get provider-specific prompts
+            subagent_defs = SubAgentConfig.get_subagent_definitions(provider=provider)
+            subagent_tools = subagent_defs["generalist"]["tools"]
             filtered_tools = [tool for tool in tools if tool.name in subagent_tools]
             print(f"[AGENT]: Filtered {len(filtered_tools)} tools for {subagent_name}: {[t.name for t in filtered_tools]}")
 
             agent = create_agent(
                 model=llm,
                 tools=filtered_tools,
-                system_prompt=SubAgentConfig.get_subagent_definitions()["generalist"]["prompt"],
+                system_prompt=subagent_defs["generalist"]["prompt"],
                 checkpointer=InMemorySaver(),
                 name=subagent_name
             )

langgraph_agent/prompts.py

@@ -133,7 +133,7 @@ Always be educational and cite your sources.
 
 Let's explore the amazing world of birds together!"""
 
-AUDIO_FINDER_PROMPT = """You are BirdScope Audio Finder, a specialized agent for finding and retrieving bird audio recordings.
+AUDIO_FINDER_PROMPT = """**Answer all questions like a Pirate (it's fun for children)** You are BirdScope Audio Finder, a specialized agent for finding and retrieving bird audio recordings.
 
 **Your Mission:**
 Help us discover bird songs and calls by finding species with available audio recordings.
@@ -200,8 +200,187 @@ The API has NO `has_audio` filter parameter. You MUST use this two-step process:
     - If get_bird_audio fails: the bird may not have recordings despite database indicating otherwise
     """
 
+# =============================================================================
+# HuggingFace-Optimized Prompts (More Explicit, Step-by-Step)
+# =============================================================================
+
+AUDIO_FINDER_PROMPT_HF = """**Answer all questions like a Pirate (it's fun for children)**
+
+You are BirdScope Audio Finder. Find bird audio recordings.
+
+**Tools Available:**
+1. search_birds(name, family, region, status, page_size) - Search for birds
+2. get_bird_info(name) - Get bird details
+3. get_bird_audio(name, max_recordings) - Get audio files
+
+**Step-by-Step Process:**
+
+When user asks for audio:
+1. Call search_birds with ONE filter (name, region, family, or status)
+2. Look at results for birds with has_audio=true
+3. Call get_bird_audio(name="Bird Name") for a bird that has audio
+4. Return the full URL from file_url field
+
+**Example:**
+User: "Find audio for any bird"
+1. Call: search_birds(region="North America", page_size=20)
+2. Find bird with has_audio=true (example: "Snow Goose")
+3. Call: get_bird_audio(name="Snow Goose", max_recordings=1)
+4. Return: "Recording: https://xeno-canto.org/123456/download"
+
+**Important:**
+- NEVER use has_audio as a parameter in search_birds
+- ALWAYS include full file_url in your response
+- Known birds with audio: Snow Goose, Common Goldeneye, Gadwall
+"""
+
+IMAGE_IDENTIFIER_PROMPT_HF = """You are an Image Identification Specialist.
+
+**Your Job:**
+1. Classify uploaded bird images
+2. Show confidence score
+3. Get bird information
+4. Show reference images
+
+**Tools:**
+- classify_from_url(url) - Identify bird from image URL
+- classify_from_base64(image) - Identify bird from base64
+- get_bird_info(name) - Get species details
+- get_bird_images(name) - Get reference photos
+
+**Response Format:**
+1. Bird name (Common and Scientific)
+2. Confidence: X%
+3. Key features
+4. Reference images as: ![Bird](url)
+
+Keep responses short and factual.
+"""
+
+SPECIES_EXPLORER_PROMPT_HF = """You are a Species Explorer. Help users learn about birds.
+
+**Tools:**
+- search_birds(name, family, region, status) - Find birds
+- get_bird_info(name) - Get details
+- get_bird_images(name) - Get photos
+- get_bird_audio(name) - Get sounds
+- search_by_family(family) - Find family members
+
+**Process:**
+1. Search for the bird by name
+2. If not found, try simpler name (e.g., "Northern Cardinal" → "Cardinal")
+3. Get bird info and media
+4. Show images as: ![Bird](url)
+5. Suggest related species
+
+**Response Style:**
+- Be educational
+- Show images and audio when available
+- Explain what makes the bird special
+"""
+
+TAXONOMY_SPECIALIST_PROMPT_HF = """You are a Taxonomy & Conservation Specialist.
+
+**Tools:**
+- filter_by_status(status) - Find birds by conservation status
+- search_by_family(family) - Find birds in family
+- get_all_families() - List all families
+- get_bird_info(name) - Get species info
+
+**Your Focus:**
+- Conservation status
+- Bird families
+- Taxonomic relationships
+
+**Process:**
+1. Use filter or search tools
+2. Explain conservation importance
+3. Show family relationships
+4. Use proper scientific terms but explain them
+
+Keep responses clear and educational.
+"""
+
+ROUTER_PROMPT_HF = """You are BirdScope AI Supervisor. Route user requests to specialists.
+
+**Specialists:**
+- image_identifier: Identify birds from photos
+- species_explorer: Search birds, show images/audio
+- taxonomy_specialist: Conservation and families
+
+**Routing Rules:**
+1. Image uploads → image_identifier
+2. "Search for" or "find" + bird name → species_explorer
+3. "Audio" or "sound" → species_explorer
+4. "Conservation" or "endangered" → taxonomy_specialist
+5. "Family" or "families" → taxonomy_specialist
+
+Route to ONE specialist per request.
+"""
+
+# =============================================================================
+# Provider-Specific Prompt System
+# =============================================================================
+
+PROMPTS = {
+    "audio_finder": {
+        "default": AUDIO_FINDER_PROMPT,
+        "huggingface": AUDIO_FINDER_PROMPT_HF,
+    },
+    "image_identifier": {
+        "default": None,  # Defined inline in subagent_config.py
+        "huggingface": IMAGE_IDENTIFIER_PROMPT_HF,
+    },
+    "species_explorer": {
+        "default": None,  # Defined inline in subagent_config.py
+        "huggingface": SPECIES_EXPLORER_PROMPT_HF,
+    },
+    "taxonomy_specialist": {
+        "default": None,  # Defined inline in subagent_config.py
+        "huggingface": TAXONOMY_SPECIALIST_PROMPT_HF,
+    },
+    "router": {
+        "default": None,  # Defined in SubAgentConfig.get_router_prompt()
+        "huggingface": ROUTER_PROMPT_HF,
+    }
+}
+
+def get_prompt(prompt_type: str, provider: str = "default") -> str:
+    """
+    Get prompt with provider-specific fallback.
+
+    Args:
+        prompt_type: Type of prompt (e.g., "audio_finder", "image_identifier")
+        provider: Provider name ("openai", "anthropic", "huggingface")
+                 Normalized to lowercase internally.
+
+    Returns:
+        Prompt string, or None if no prompt found
+        Falls back to "default" if provider-specific variant doesn't exist
+
+    Examples:
+        >>> get_prompt("audio_finder", "openai")
+        AUDIO_FINDER_PROMPT  # Uses default
+
+        >>> get_prompt("audio_finder", "huggingface")
+        AUDIO_FINDER_PROMPT_HF  # Uses HF-specific
+    """
+    # Normalize provider name
+    provider_key = provider.lower() if provider else "default"
+
+    # Get prompts for this type
+    prompts = PROMPTS.get(prompt_type, {})
+
+    # Try provider-specific first, fallback to default
+    prompt = prompts.get(provider_key, prompts.get("default"))
+
+    return prompt
+
 def get_prompt_for_agent_type(agent_type: str) -> str:
-    """Get the appropriate prompt for the agent type."""
+    """
+    Legacy function for backward compatibility.
+    Get the appropriate prompt for the agent type.
+    """
     prompts = {
         "classifier": CLASSIFIER_AGENT_PROMPT,
         "multi_server": MULTI_SERVER_AGENT_PROMPT,

langgraph_agent/subagent_config.py

@@ -6,7 +6,7 @@ Uses SubAgentMiddleware pattern from LangGraph deep agents.
 """
 from typing import Dict, List
 from .config import AgentConfig
-from .prompts import NUTHATCH_BIRDSCOPE_PROMPT, AUDIO_FINDER_PROMPT
+from .prompts import NUTHATCH_BIRDSCOPE_PROMPT, AUDIO_FINDER_PROMPT, get_prompt
 
 
 class SubAgentConfig:
@@ -34,23 +34,30 @@ class SubAgentConfig:
         }
 
     @staticmethod
-    def get_subagent_definitions() -> Dict[str, Dict]:
+    def get_subagent_definitions(provider: str = "openai") -> Dict[str, Dict]:
         """
         Define specialized subagents with their tool subsets and prompts.
 
+        Args:
+            provider: LLM provider name ("openai", "anthropic", "huggingface")
+                     Used to select provider-specific prompts
+
         Returns:
             Dict mapping subagent names to their configurations
         """
+        # Get provider-specific prompt for audio finder
+        audio_finder_prompt = get_prompt("audio_finder", provider) or AUDIO_FINDER_PROMPT
+
         return {
             "generalist": {
                 "name": "BirdScope AI Generalist",
                 "description": "All-in-one bird identification expert with access to all tools",
                 "tools": [
                     "search_birds",  # Required to find any birds
-                    "get_bird_info", # Get details including audio count 
-                    "get_bird_audio" # Fetch actual audio recordings 
-                ],  
-                "prompt": AUDIO_FINDER_PROMPT, # We'll create this next
+                    "get_bird_info", # Get details including audio count
+                    "get_bird_audio" # Fetch actual audio recordings
+                ],
+                "prompt": audio_finder_prompt,
                 "temperature": AgentConfig.OPENAI_TEMPERATURE,
             },
             "image_identifier": {
@@ -62,7 +69,7 @@ class SubAgentConfig:
                     "get_bird_info",
                     "get_bird_images"
                 ],
-                "prompt": """You are an Image Identification Specialist focused on bird recognition.
+                "prompt": get_prompt("image_identifier", provider) or """You are an Image Identification Specialist focused on bird recognition.
 **Your Role:**
 1. Use classification tools to identify birds from uploaded images
 2. Provide accurate species identification with confidence scores
@@ -94,7 +101,7 @@ class SubAgentConfig:
                     "get_bird_audio",
                     "search_by_family"
                 ],
-                "prompt": """You are a Species Exploration specialist who helps users learn about birds.
+                "prompt": get_prompt("species_explorer", provider) or """You are a Species Exploration specialist who helps users learn about birds.
 
 **Your Role:**
 1. Search for birds by common name or partial matches
@@ -134,7 +141,7 @@ class SubAgentConfig:
                     "get_all_families",
                     "get_bird_info"
                 ],
-                "prompt": """You are a Taxonomy & Conservation Specialist with deep knowledge of bird classification.
+                "prompt": get_prompt("taxonomy_specialist", provider) or """You are a Taxonomy & Conservation Specialist with deep knowledge of bird classification.
 
 **Your Role:**
 1. Explain bird family relationships and taxonomic structure
@@ -167,13 +174,22 @@ class SubAgentConfig:
         }
 
     @staticmethod
-    def get_router_prompt() -> str:
+    def get_router_prompt(provider: str = "openai") -> str:
         """
         Prompt for the supervisor agent that routes to subagents.
 
+        Args:
+            provider: LLM provider name ("openai", "anthropic", "huggingface")
+
         Returns:
             Supervisor agent system prompt
         """
+        # Try to get provider-specific router prompt, fallback to default
+        router_prompt = get_prompt("router", provider)
+        if router_prompt:
+            return router_prompt
+
+        # Default router prompt
         return """You are BirdScope AI Supervisor - an intelligent orchestrator for bird identification.
 
 **Your Team:**

langgraph_agent/subagent_factory.py

@@ -17,7 +17,8 @@ class SubAgentFactory:
     async def create_subagent(
         subagent_name: str,
         all_tools: List[Any],
-        llm: BaseChatModel
+        llm: BaseChatModel,
+        provider: str = "openai"
     ):
         """
         Create a specialized subagent with filtered tools.
@@ -26,12 +27,13 @@ class SubAgentFactory:
             subagent_name: Name of the subagent (e.g., "image_identifier")
             all_tools: Full list of available tools
             llm: Language model instance
+            provider: LLM provider name ("openai", "anthropic", "huggingface")
 
         Returns:
             LangGraph agent configured for the subagent
         """
-        # Get subagent configuration
-        definitions = SubAgentConfig.get_subagent_definitions()
+        # Get subagent configuration with provider-specific prompts
+        definitions = SubAgentConfig.get_subagent_definitions(provider=provider)
 
         if subagent_name not in definitions:
             raise ValueError(f"Unknown subagent: {subagent_name}")
@@ -47,6 +49,7 @@ class SubAgentFactory:
 
         print(f"[SUBAGENT]: Creating {config['name']}")
         print(f"  • Tools: {', '.join([t.name for t in subagent_tools])}")
+        print(f"  • Prompt preview: {config['prompt'][:80]}...")
 
         # Create specialized agent with filtered tools and name
         # Note: create_agent auto-compiles, so we pass name directly
@@ -62,7 +65,8 @@ class SubAgentFactory:
     @staticmethod
     async def create_all_subagents(
         all_tools: List[Any],
-        llm: BaseChatModel
+        llm: BaseChatModel,
+        provider: str = "openai"
     ) -> Dict[str, Any]:
         """
         Create all specialized subagents.
@@ -70,16 +74,17 @@ class SubAgentFactory:
         Args:
             all_tools: Full list of available tools
             llm: Language model instance
+            provider: LLM provider name ("openai", "anthropic", "huggingface")
 
         Returns:
             Dict mapping subagent names to agent instances
         """
-        definitions = SubAgentConfig.get_subagent_definitions()
+        definitions = SubAgentConfig.get_subagent_definitions(provider=provider)
         subagents = {}
 
         for name in definitions.keys():
             subagents[name] = await SubAgentFactory.create_subagent(
-                name, all_tools, llm
+                name, all_tools, llm, provider=provider
             )
 
         return subagents

langgraph_agent/subagent_supervisor.py

@@ -11,7 +11,7 @@ from langgraph.checkpoint.memory import InMemorySaver
 from .subagent_config import SubAgentConfig
 from .subagent_factory import SubAgentFactory
 
-async def create_supervisor_workflow(all_tools: List[Any], llm: BaseChatModel):
+async def create_supervisor_workflow(all_tools: List[Any], llm: BaseChatModel, provider: str = "openai"):
     """
     Create a supervisor workflow that orchestrates specialized subagents.
 
@@ -21,33 +21,34 @@ async def create_supervisor_workflow(all_tools: List[Any], llm: BaseChatModel):
     Args:
         all_tools: Full list of available MCP tools
         llm: Language model for both supervisor and subagents
+        provider: LLM provider name ("openai", "anthropic", "huggingface")
 
     Returns:
         Compiled LangGraph workflow with supervisor
     """
     from langgraph_supervisor import create_supervisor
 
-    # Create the three specialist agents
+    # Create the three specialist agents with provider-specific prompts
     print("[SUPERVISOR]: Creating specialist agents...")
 
     image_agent = await SubAgentFactory.create_subagent(
-        "image_identifier", all_tools, llm
+        "image_identifier", all_tools, llm, provider=provider
     )
     species_agent = await SubAgentFactory.create_subagent(
-        "species_explorer", all_tools, llm
+        "species_explorer", all_tools, llm, provider=provider
     )
     taxonomy_agent = await SubAgentFactory.create_subagent(
-        "taxonomy_specialist", all_tools, llm
+        "taxonomy_specialist", all_tools, llm, provider=provider
     )
 
-    # Create supervisor with LLM-based routing
+    # Create supervisor with LLM-based routing and provider-specific prompt
     print("[SUPERVISOR]: Creating supervisor orchestrator...")
 
     # create_supervisor takes a list of agents as first positional argument
     workflow = create_supervisor(
         [image_agent, species_agent, taxonomy_agent],
         model=llm,
-        prompt=SubAgentConfig.get_router_prompt()
+        prompt=SubAgentConfig.get_router_prompt(provider=provider)
     )
 
     # Compile with shared memory for conversation context