From b9311036d03dbf2956d5506ac371ff4bae6f25b1 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Erik=20Bj=C3=A4reholt?= <erik@bjareho.lt>
Date: Wed, 30 Oct 2024 13:55:13 +0100
Subject: [PATCH] fix: improved reliability of llm.generate_name()

---
 gptme/commands.py | 13 +++++++++----
 gptme/llm.py      | 14 ++++++++++++--
 2 files changed, 21 insertions(+), 6 deletions(-)

diff --git a/gptme/commands.py b/gptme/commands.py
index 139e00a6..78a6d97e 100644
--- a/gptme/commands.py
+++ b/gptme/commands.py
@@ -89,8 +89,12 @@ def handle_cmd(
             manager.undo(1, quiet=True)
             manager.write()
             # rename the conversation
-            print("Renaming conversation (enter empty name to auto-generate)")
-            new_name = args[0] if args else input("New name: ")
+            print("Renaming conversation")
+            if args:
+                new_name = args[0]
+            else:
+                print("(enter empty name to auto-generate)")
+                new_name = input("New name: ").strip()
             rename(manager, new_name, confirm)
         case "fork":
             # fork the conversation
@@ -193,8 +197,9 @@ def edit(manager: LogManager) -> Generator[Message, None, None]:  # pragma: no c
 
 def rename(manager: LogManager, new_name: str, confirm: ConfirmFunc) -> None:
     if new_name in ["", "auto"]:
-        new_name = llm.generate_name(prepare_messages(manager.log.messages))
-        assert " " not in new_name
+        msgs = prepare_messages(manager.log.messages)[1:]  # skip system message
+        new_name = llm.generate_name(msgs)
+        assert " " not in new_name, f"Invalid name: {new_name}"
         print(f"Generated name: {new_name}")
         if not confirm("Confirm?"):
             print("Aborting")
diff --git a/gptme/llm.py b/gptme/llm.py
index cc738cd5..99407480 100644
--- a/gptme/llm.py
+++ b/gptme/llm.py
@@ -164,12 +164,17 @@ def generate_name(msgs: list[Message]) -> str:
     """
     # filter out system messages
     msgs = [m for m in msgs if m.role != "system"]
+
+    # TODO: filter out assistant messages? (only for long conversations? or always?)
+    # msgs = [m for m in msgs if m.role != "assistant"]
+
     msgs = (
         [
             Message(
                 "system",
                 """
-The following is a conversation between a user and an assistant. Which we will generate a name for.
+The following is a conversation between a user and an assistant.
+You should generate a descriptive name for it.
 
 The name should be 3-6 words describing the conversation, separated by dashes. Examples:
  - install-llama
@@ -183,7 +188,12 @@ def generate_name(msgs: list[Message]) -> str:
             )
         ]
         + msgs
-        + [Message("user", "Now, generate a name for this conversation.")]
+        + [
+            Message(
+                "user",
+                "That was the context of the conversation. Now, answer with a descriptive name for this conversation according to system instructions.",
+            )
+        ]
     )
     name = _chat_complete(msgs, model=get_summary_model(_client_to_provider())).strip()
     return name