提交: 3c803c8de0d04c3fb45e9e0455f8ab2111e59206 - zirpu (git)

Emergent generative agents

Commit MetaInfo

• 修訂: 3c803c8de0d04c3fb45e9e0455f8ab2111e59206 (tree)
• 時間: 2023-04-15 00:16:39
• 作者: Corbin <cds@corb...>
• Commiter: Corbin

Log Message

Fix some newlines, and experiment with thoughts.

Trying to avoid picking exponentially-long chains of thought by
preferring shorter thoughts. First, select relevant thoughts; then,
select shortest thoughts.

Change Summary

• modified: agent.py
• modified: append_thought.py
• modified: common.py

差異

--- a/agent.py
+++ b/agent.py
@@ -3,6 +3,7 @@
 
 from concurrent.futures import ThreadPoolExecutor
 from datetime import datetime
+from heapq import nsmallest
 import json
 import os.path
 import random
@@ -37,10 +38,10 @@ def build_prompt(persona, shadow, flavorText="", selfAware=False,
     return f"""Character Persona: {titledName} ({personaTraits})
 
 Character Shadow Persona: {shadow[0]} ({shadowTraits})
-
-Flavor text: {flavorText}
 """, titledName
 
+# Flavor text: {flavorText}
+
 def load_character(path):
     with open(os.path.join(path, "character.json"), "r") as handle:
         return json.load(handle)
@@ -60,7 +61,7 @@ class SentenceIndex:
         with open(self.path, "w") as f: json.dump(dict(self.db), f)
 
     def search(self, embedding, k):
-        with Timer("k nearest neighbors"):
+        with Timer("%d nearest neighbors" % k):
             D, I = self.index.search(np.array([embedding], dtype="float32"), k)
             return [self.db[i][0] for i in I[0] if i >= 0]
 
@@ -140,7 +141,12 @@ class Agent(SingleServerIRCBot):
 
     def thoughtPrompt(self):
         key = NO_THOUGHTS_EMBED if self.recent_thought is None else self.recent_thought[1]
-        new_thoughts = thought_index.search(key, 4)
+        # Fetch more thoughts than necessary, and always prefer shorter
+        # thoughts. This is an attempt to prevent exponential rumination.
+        new_thoughts = thought_index.search(key, 10)
+        # .search() returns most relevant thoughts first; reversing the list
+        # creates more focused chains of thought.
+        new_thoughts = nsmallest(5, new_thoughts.reverse(), key=len)
         if self.recent_thought is not None:
             new_thoughts.append(self.recent_thought[0])
         print("~ Thoughts:", *new_thoughts)
@@ -148,8 +154,9 @@ class Agent(SingleServerIRCBot):
 
     def newThoughtPrompt(self, channel):
         lines = self.logs[channel].l[-10:]
-        return (self.thoughtPrompt() + self.chatPrompt(channel) +
-                "\n".join(lines) + "\nEND OF LOG\n")
+        return (self.chatPrompt(channel) + "\n" +
+                "\n".join(lines) + "\nEND OF LOG\n" +
+                self.thoughtPrompt())
 
     def chatPrompt(self, channel):
         c = self.channels[channel]
@@ -178,7 +185,7 @@ Users: {users}"""
         prefix = f"{datetime.now():%H:%M:%S} <{nick}>"
         examples = self.examplesFromOtherChannels(channel)
         # NB: "full" prompt needs log lines from current channel...
-        fullPrompt = prompt + self.thoughtPrompt() + examples + self.chatPrompt(channel)
+        fullPrompt = prompt + self.thoughtPrompt() + examples + self.chatPrompt(channel) + "\n"
         # ...but we need to adjust the log offset first...
         log.bumpCutoff(max_context_length, gen.countTokens, fullPrompt, prefix)
         # ...and current channel's log lines are added here.
@@ -196,9 +203,7 @@ Users: {users}"""
 
     def thinkAbout(self, channel):
         print("~ Will ponder channel:", channel)
-        thoughtPrompt = prompt + self.newThoughtPrompt(channel)
-        prefix = "New thought:"
-        s = thoughtPrompt + prefix
+        s = prompt + self.newThoughtPrompt(channel)
         def cb(completion):
             self.thinking = False
             thought = breakIRCLine(completion.result())

--- a/append_thought.py
+++ b/append_thought.py
@@ -9,6 +9,7 @@ path = sys.argv[-1]
 gen = CamelidGen()
 
 with open(path, "r") as handle: db = json.load(handle)
+print("Thought database:", len(db), "entries")
 
 while True:
     try: thought = input("> ").strip()
@@ -16,4 +17,5 @@ while True:
     if not thought: break
     db[thought] = gen.embed(thought)
 
+print("Saving thought database:", len(db), "entries")
 with open(path, "w") as handle: json.dump(db, handle)

--- a/common.py
+++ b/common.py
@@ -29,7 +29,7 @@ class Log:
         return self.finishPromptAtCutoff(self.cutoff, s, prefix)
 
     def finishPromptAtCutoff(self, cutoff, s, prefix):
-        return s + "\n".join(self.l[cutoff:]) + prefix
+        return s + "\n".join(self.l[cutoff:]) + "\n" + prefix
 
     def undo(self): self.l.pop()