properly feed in conversation history

.containerignore

@@ -2,3 +2,6 @@
 .mypy_cache
 .vscode
 .token
+*.db
+*.env
+*.png

Containerfile

@@ -2,6 +2,9 @@ FROM ghcr.io/astral-sh/uv:python3.13-alpine
 
 # Copy the project into the image
 COPY vibe_bot /app
+COPY uv.lock /app
+COPY .python-version /app
+COPY pyproject.toml /app
 
 # Disable development dependencies
 ENV UV_NO_DEV=1

README.md

@@ -188,8 +188,7 @@ uv run main.py
 podman build -t vibe-bot:latest .
 
 # Run
-export DISCORD_TOKEN=$(cat .token)
+podman run --env-file .env localhost/vibe-bot:latest
-podman run -e DISCORD_TOKEN localhost/vibe-bot:latest
 ```
 
 ## Docs

vibe_bot/database.py

@@ -346,7 +346,7 @@ class ChatDatabase:
 
     def get_conversation_context(
         self, user_id: str, current_message: str, max_context: int = 5
-    ) -> str:
+    ) -> list[dict[str, str]]:
         """Get relevant conversation context for RAG."""
         # Get recent messages from the user
         recent_messages = self.get_user_history(user_id, limit=max_context * 2)
@@ -357,22 +357,21 @@ class ChatDatabase:
         )
 
         # Combine contexts
-        context_parts = []
+        context_parts: list[dict[str, str]] = []
 
         # Add recent messages
         for user_message, bot_message in recent_messages:
-            combined_content = f"[Recent chat]\n{user_message}\n{bot_message}"
+            context_parts.append({"role": "assistant", "content": bot_message})
-            context_parts.append(combined_content)
+            context_parts.append({"role": "user", "content": user_message})
 
         # Add similar messages
         for user_message, bot_message, similarity in similar_messages:
-            combined_content = f"{user_message}\n{bot_message}"
+            context_parts.append({"role": "assistant", "content": bot_message})
-            if combined_content not in "\n".join(context_parts):
+            context_parts.append({"role": "user", "content": user_message})
-                context_parts.append(f"[You remember]\n{combined_content}")
 
         # Conversation history needs to be delivered in "newest context last" order
         context_parts.reverse()
-        return "\n".join(context_parts[-max_context * 4 :])  # Limit total context
+        return context_parts
 
     def clear_all_messages(self):
         """Clear all messages and embeddings from the database."""

vibe_bot/llama_wrapper.py

@@ -44,6 +44,42 @@ def chat_completion(
         return ""
 
 
+def chat_completion_with_history(
+    system_prompt: str,
+    prompts: Iterable[ChatCompletionMessageParam],
+    openai_url: str,
+    openai_api_key: str,
+    model: str,
+    max_tokens: int = 1000,
+) -> str:
+    client = openai.OpenAI(base_url=openai_url, api_key=openai_api_key)
+    messages: Iterable[ChatCompletionMessageParam] = [
+        {
+            "role": "system",
+            "content": system_prompt,
+        }
+    ] + prompts  # type: ignore
+    response = client.chat.completions.create(
+        model=model,
+        messages=messages,
+        max_tokens=max_tokens,
+        extra_body={
+            "chat_template_kwargs": {"enable_thinking": False},
+        },
+        seed=-1,
+    )
+
+    # Assert that thinking was used
+    if response.choices[0].message.model_extra:
+        assert response.choices[0].message.model_extra.get("reasoning_content")
+
+    content = response.choices[0].message.content
+    if content:
+        return content.strip()
+    else:
+        return ""
+
+
 def chat_completion_instruct(
     system_prompt: str,
     user_prompt: str,

vibe_bot/main.py

@@ -297,6 +297,112 @@ async def retcon(ctx, *, message: str):
     await ctx.send(file=send_img)
 
 
+@bot.command(name="talkforme")
+async def talkforme(ctx, *, message: str):
+    """Have two bots talk to each other about a topic
+
+    Usage: !talkforme bot1 bot2 4 some conversation topic
+    """
+
+    TALK_LIMIT = 20
+
+    bot1_name, bot2_name, limit, topic_list = (
+        message.split(" ")[0],
+        message.split(" ")[1],
+        message.split(" ")[2],
+        message.split(" ")[3:],
+    )
+
+    topic = " ".join(topic_list)
+
+    custom_bot_manager = CustomBotManager()
+    bot1 = custom_bot_manager.get_custom_bot(bot1_name)
+
+    if not bot1:
+        await ctx.send(f"{bot1_name} is not a real bot...")
+        return
+    else:
+        _, bot1_prompt, _, _ = bot1
+
+    bot2 = custom_bot_manager.get_custom_bot(bot2_name)
+
+    if not bot2:
+        await ctx.send(f"{bot2_name} is not a real bot...")
+        return
+    else:
+        _, bot2_prompt, _, _ = bot2
+
+    await ctx.send(
+        f'{bot1_name} is going to talk to {bot2_name} about "{topic[:50]}" for {limit} replies.'
+    )
+
+    bot_list = [(bot1_name, bot1_prompt), (bot2_name, bot2_prompt)]
+
+    message_limit = int(limit)
+
+    def flip_counter(counter: int):
+        if counter == 0:
+            return 1
+        else:
+            return 0
+
+    def flip_user(user: str):
+        if user == "user":
+            return "assistant"
+        else:
+            return "user"
+
+    message_counter = 0
+    bot_counter = 0
+    current_bot = bot_list[bot_counter]
+    prompt_histories = [
+        [{"role": "user", "content": topic}],
+        [{"role": "assistant", "content": topic}],
+    ]
+
+    first_bot_response = llama_wrapper.chat_completion_with_history(
+        system_prompt=current_bot[1]
+        + f"\nKeep your responses under 2-3 sentences. You are talking to {current_bot[flip_counter(bot_counter)][0]}",
+        prompts=prompt_histories[bot_counter],  # type: ignore
+        openai_url=CHAT_ENDPOINT,
+        openai_api_key=CHAT_ENDPOINT_KEY,
+        model=CHAT_MODEL,
+        max_tokens=MAX_COMPLETION_TOKENS,
+    )
+    await ctx.send(f"## {current_bot[0]}\n{first_bot_response}")
+    prompt_histories[0].append({"role": "assistant", "content": first_bot_response})
+    prompt_histories[1].append({"role": "user", "content": first_bot_response})
+
+    bot_counter = flip_counter(counter=bot_counter)
+
+    while message_counter < min(message_limit, TALK_LIMIT):
+        current_bot = bot_list[bot_counter]
+        logger.info(f"Current bot is {current_bot}")
+        bot_response = llama_wrapper.chat_completion_with_history(
+            system_prompt=current_bot[1]
+            + f"\nKeep your responses under 2-3 sentences. {current_bot[flip_counter(bot_counter)]}",
+            prompts=prompt_histories[bot_counter],  # type: ignore
+            openai_url=CHAT_ENDPOINT,
+            openai_api_key=CHAT_ENDPOINT_KEY,
+            model=CHAT_MODEL,
+            max_tokens=MAX_COMPLETION_TOKENS,
+        )
+        message_counter += 1
+        prompt_histories[bot_counter].append(
+            {"role": "assistant", "content": bot_response}
+        )
+        prompt_histories[flip_counter(bot_counter)].append(
+            {"role": "user", "content": bot_response}
+        )
+        await ctx.send(f"## {current_bot[0]}")
+        while bot_response:
+            send_chunk = bot_response[:1000]
+            bot_response = bot_response[1000:]
+            await ctx.send(send_chunk)
+        bot_counter = flip_counter(counter=bot_counter)
+        logger.info(f"Message counter is {message_counter}/{limit}")
+
+
 async def handle_chat(
     ctx, *, bot_name: str, message: str, system_prompt: str, response_prefix: str
 ):
@@ -310,22 +416,19 @@ async def handle_chat(
         user_id=str(ctx.author.id), current_message=message, max_context=5
     )
 
+    prompts = [{"role": "user", "content": message}]
+
     if context:
-        user_message = f"\n\nRelevant conversation history:\n{context}\n\n{message}"
+        prompts = context + prompts
-    else:
-        user_message = message
 
-    logger.info(user_message)
+    logger.info(prompts)
 
-    system_prompt_edit = (
+    system_prompt_edit = f"{system_prompt}\nKeep your responses under 2-3 sentences."
-        "Keep your responses somewhat short, limited to 500 words or less. "
-        f"{system_prompt}"
-    )
 
     try:
-        bot_response = llama_wrapper.chat_completion_instruct(
+        bot_response = llama_wrapper.chat_completion_with_history(
             system_prompt=system_prompt_edit,
-            user_prompt=user_message,
+            prompts=prompts,  # type: ignore
             openai_url=CHAT_ENDPOINT,
             openai_api_key=CHAT_ENDPOINT_KEY,
             model=CHAT_MODEL,