-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy path53_prompt_compression.py
More file actions
263 lines (212 loc) · 8.75 KB
/
53_prompt_compression.py
File metadata and controls
263 lines (212 loc) · 8.75 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
#!/usr/bin/env python3
"""
Prompt Compression — prevent context-window overflow in long conversations.
Demonstrates:
- compress_context=True — enable proactive context compression
- compress_threshold=0.75 — trigger when context is 75 % full
- compress_keep_recent=4 — keep last N turns verbatim
- PROMPT_COMPRESSED step in AgentTrace
- Observer event: on_prompt_compressed(run_id, before_tokens, after_tokens, count)
- Memory vs history: self.memory is never modified; only the per-call view
How it works:
Before each LLM call, selectools estimates the token count. If the fill
rate exceeds compress_threshold, older messages are summarised into a single
[Compressed context] system message. Recent turns (compress_keep_recent)
are always kept verbatim so the LLM retains immediate context.
Prerequisites:
pip install selectools
# Uses a stub provider — no API keys needed.
"""
from __future__ import annotations
from typing import List, Optional, Tuple
from unittest.mock import patch
from selectools import Agent, AgentConfig, Message, Role, Tool, UsageStats
from selectools.memory import ConversationMemory
from selectools.observer import AgentObserver
from selectools.token_estimation import TokenEstimate
from selectools.trace import StepType
# ---------------------------------------------------------------------------
# Stub provider (no real API calls)
# ---------------------------------------------------------------------------
class StubProvider:
"""Returns scripted responses in order."""
name = "stub"
supports_streaming = False
supports_async = False
def __init__(self, responses: List[str]) -> None:
self._responses = responses
self._idx = 0
def complete(
self,
*,
model: str,
system_prompt: str,
messages: List[Message],
tools: Optional[List[Tool]] = None,
temperature: float = 0.0,
max_tokens: int = 1000,
timeout: Optional[float] = None,
) -> Tuple[Message, UsageStats]:
text = self._responses[min(self._idx, len(self._responses) - 1)]
self._idx += 1
usage = UsageStats(
prompt_tokens=100,
completion_tokens=20,
total_tokens=120,
cost_usd=0.0,
model=model,
provider="stub",
)
return Message(role=Role.ASSISTANT, content=text), usage
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def _noop_tool() -> Tool:
return Tool(name="noop", description="no-op", parameters=[], function=lambda: "ok")
def _make_estimate(total_tokens: int) -> TokenEstimate:
ctx = 100_000
return TokenEstimate(
0, total_tokens, 0, total_tokens, ctx, ctx - total_tokens, "stub", "heuristic"
)
def _fake_history(n_turns: int) -> List[Message]:
"""n_turns user/assistant pairs."""
msgs: List[Message] = []
for i in range(n_turns):
msgs.append(Message(role=Role.USER, content=f"Tell me about topic {i}."))
msgs.append(Message(role=Role.ASSISTANT, content=f"Topic {i} is about X, Y, and Z."))
return msgs
class CompressionObserver(AgentObserver):
"""Records on_prompt_compressed events."""
def __init__(self) -> None:
self.events: list = []
def on_prompt_compressed(
self,
run_id: str,
before_tokens: int,
after_tokens: int,
messages_compressed: int,
) -> None:
self.events.append(
{
"before": before_tokens,
"after": after_tokens,
"compressed": messages_compressed,
}
)
print(
f" [observer] Compressed {messages_compressed} messages: "
f"{before_tokens:,} → {after_tokens:,} tokens"
)
# ---------------------------------------------------------------------------
# Demo
# ---------------------------------------------------------------------------
def _separator(title: str) -> None:
print(f"\n{'=' * 55}")
print(f" {title}")
print("=" * 55)
def main() -> None:
print("=== Prompt Compression Demo ===")
# ------------------------------------------------------------------ #
# 1. Disabled by default
# ------------------------------------------------------------------ #
_separator("1. Disabled by default")
agent = Agent(
tools=[_noop_tool()],
provider=StubProvider(["hello"]),
config=AgentConfig(model="gpt-4o-mini"), # compress_context defaults to False
)
result = agent.run("hi")
compressed = [s for s in result.trace.steps if s.type == StepType.PROMPT_COMPRESSED]
print(f" compress_context default: {agent.config.compress_context}")
print(f" PROMPT_COMPRESSED steps: {len(compressed)} (expected 0)")
# ------------------------------------------------------------------ #
# 2. Fires when context is nearly full
# ------------------------------------------------------------------ #
_separator("2. Compression fires when threshold exceeded")
observer = CompressionObserver()
provider = StubProvider(["Summary of old messages.", "Final answer."])
config = AgentConfig(
model="gpt-4o-mini",
compress_context=True,
compress_threshold=0.85, # trigger at 85 % fill
compress_keep_recent=1, # keep last 1 turn verbatim
observers=[observer],
)
agent2 = Agent(tools=[_noop_tool()], provider=provider, config=config)
# Pre-load 5 turns of history
agent2.memory = ConversationMemory(max_messages=50)
for msg in _fake_history(5):
agent2.memory.add(msg)
print(f" Memory messages before run: {len(agent2.memory)}")
print(f" Threshold: {config.compress_threshold} (85 %)")
# Patch estimate so the first call looks near-full (90 k / 100 k = 90 %)
def _toggle_estimate(*args, **kwargs):
if not hasattr(_toggle_estimate, "_called"):
_toggle_estimate._called = True
return _make_estimate(90_000)
return _make_estimate(2_000)
with patch(
"selectools.agent._memory_manager.estimate_run_tokens", side_effect=_toggle_estimate
):
result2 = agent2.run("What have we covered so far?")
compressed_steps = [s for s in result2.trace.steps if s.type == StepType.PROMPT_COMPRESSED]
print(f"\n PROMPT_COMPRESSED steps: {len(compressed_steps)}")
if compressed_steps:
step = compressed_steps[0]
print(f" Before → After: {step.prompt_tokens:,} → {step.completion_tokens:,} tokens")
print(f" Summary: {step.summary}")
print(f" Memory UNCHANGED: {len(agent2.memory)} messages (history view was compressed)")
# ------------------------------------------------------------------ #
# 3. keep_recent preserves recent turns
# ------------------------------------------------------------------ #
_separator("3. compress_keep_recent controls verbatim window")
print(
"""
config = AgentConfig(
compress_context=True,
compress_threshold=0.75, # trigger at 75 % fill
compress_keep_recent=4, # always keep last 4 turns verbatim
)
With a 10-turn history:
- Turns 0-5 → summarised into [Compressed context] system message
- Turns 6-9 → kept verbatim (last 4 turns × 2 messages each = 8 msgs)
Effect: LLM sees full recent context + condensed summary of the rest.
"""
)
# ------------------------------------------------------------------ #
# 4. Observer integration
# ------------------------------------------------------------------ #
_separator("4. Observer event: on_prompt_compressed")
print(
"""
class MyObserver(AgentObserver):
def on_prompt_compressed(
self,
run_id: str,
before_tokens: int,
after_tokens: int,
messages_compressed: int,
) -> None:
reduction = 1 - after_tokens / before_tokens
print(f"Compressed {messages_compressed} messages, "
f"{reduction:.0%} token reduction")
agent = Agent(
tools=[...],
config=AgentConfig(
compress_context=True,
observers=[MyObserver()],
),
)
"""
)
# ------------------------------------------------------------------ #
# 5. Configuration reference
# ------------------------------------------------------------------ #
_separator("5. Configuration defaults")
cfg = AgentConfig()
print(f" compress_context = {cfg.compress_context} (disabled by default)")
print(f" compress_threshold = {cfg.compress_threshold} (trigger at 75 % fill)")
print(f" compress_keep_recent = {cfg.compress_keep_recent} (keep last 4 turns verbatim)")
print("\nDone.")
if __name__ == "__main__":
main()