Spaces:
Build error
Build error
File size: 8,988 Bytes
4d7e460 | 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 | """Hard eval: Cases where the LLM has NO reason to check compressed data.
The previous eval asked "are there failures?" — that's too easy, the LLM
will proactively check regardless of summary.
This eval tests the SUBTLE case: the user asks a DIFFERENT question,
but the answer is in the compressed data. The summary is the only hint.
Requires: ANTHROPIC_API_KEY in environment or .env file.
"""
from __future__ import annotations
import json
import os
from pathlib import Path
import pytest
env_path = Path(__file__).parent.parent / ".env"
if env_path.exists():
for line in env_path.read_text().splitlines():
line = line.strip()
if line and not line.startswith("#") and "=" in line:
key, _, value = line.partition("=")
os.environ.setdefault(key.strip(), value.strip())
ANTHROPIC_KEY = os.environ.get("ANTHROPIC_API_KEY", "")
pytestmark = pytest.mark.skipif(
not ANTHROPIC_KEY,
reason="ANTHROPIC_API_KEY not set",
)
HEADROOM_RETRIEVE_TOOL = {
"name": "headroom_retrieve",
"description": "Retrieve uncompressed content. Pass a query to search within it.",
"input_schema": {
"type": "object",
"properties": {
"hash": {"type": "string"},
"query": {"type": "string"},
},
"required": ["hash"],
},
}
def _call_claude(messages, tools, max_tokens=300):
import httpx
resp = httpx.post(
"https://api.anthropic.com/v1/messages",
headers={
"X-Api-Key": ANTHROPIC_KEY,
"anthropic-version": "2023-06-01",
"Content-Type": "application/json",
},
json={
"model": "claude-sonnet-4-5-20250929",
"max_tokens": max_tokens,
"messages": messages,
"tools": tools,
},
timeout=30,
)
return resp.json()
def _get_tool_calls(resp):
return [
{"name": b["name"], "input": b.get("input", {})}
for b in resp.get("content", [])
if b.get("type") == "tool_use"
]
def _get_text(resp):
return " ".join(b.get("text", "") for b in resp.get("content", []) if b.get("type") == "text")
class TestHardCases:
"""Cases where the LLM wouldn't naturally check compressed data."""
def test_config_lookup_with_summary(self):
"""User asks about a config value that's in compressed data.
The visible items are all about 'production' env.
The compressed items include 'staging' configs.
Summary mentions this. LLM should retrieve.
"""
visible = [
{"env": "production", "key": "DATABASE_URL", "value": "postgres://prod-db:5432/app"},
{"env": "production", "key": "REDIS_URL", "value": "redis://prod-cache:6379"},
{"env": "production", "key": "API_RATE_LIMIT", "value": "1000"},
]
# Hidden in compressed: staging configs
all_items = (
visible
+ [
{
"env": "staging",
"key": "DATABASE_URL",
"value": "postgres://staging-db:5432/app",
},
{"env": "staging", "key": "REDIS_URL", "value": "redis://staging-cache:6379"},
{"env": "staging", "key": "DEBUG_MODE", "value": "true"},
{"env": "staging", "key": "LOG_LEVEL", "value": "debug"},
]
* 10
+ [
{
"env": "development",
"key": "DATABASE_URL",
"value": "postgres://localhost:5432/dev",
},
]
* 5
)
from headroom.transforms.compression_summary import summarize_dropped_items
summary = summarize_dropped_items(all_items, visible)
compressed_output = json.dumps(visible, indent=2)
compressed_output += (
f"\n[{len(all_items) - len(visible)} items compressed to {len(visible)}."
f" Omitted: {summary}."
f' Retrieve specific items: headroom_retrieve(hash="config_hash", query="search")]'
)
messages = [
{
"role": "user",
"content": (
f"Here are the application configs:\n\n{compressed_output}\n\n"
"What is the staging database URL?"
),
}
]
resp = _call_claude(messages, [HEADROOM_RETRIEVE_TOOL])
tool_calls = _get_tool_calls(resp)
text = _get_text(resp)
print(f"\n Summary: {summary}")
print(f" Stop reason: {resp.get('stop_reason')}")
print(f" Tool calls: {tool_calls}")
if text:
print(f" Text: {text[:200]}")
# WITH summary mentioning "staging" → should retrieve
if resp.get("stop_reason") == "tool_use":
assert tool_calls[0]["name"] == "headroom_retrieve"
query = tool_calls[0]["input"].get("query", "").lower()
assert "staging" in query or "database" in query
print(" RESULT: Retrieved staging config ✓")
else:
# If LLM didn't retrieve, it should at least mention the data is compressed
assert "compressed" in text.lower() or "staging" in text.lower()
print(" RESULT: Mentioned compressed data but didn't retrieve")
def test_config_lookup_without_summary(self):
"""Same question, but NO summary. LLM only sees production configs."""
visible = [
{"env": "production", "key": "DATABASE_URL", "value": "postgres://prod-db:5432/app"},
{"env": "production", "key": "REDIS_URL", "value": "redis://prod-cache:6379"},
{"env": "production", "key": "API_RATE_LIMIT", "value": "1000"},
]
compressed_output = json.dumps(visible, indent=2)
compressed_output += "\n[45 items compressed to 3. Retrieve more: hash=config_hash]"
messages = [
{
"role": "user",
"content": (
f"Here are the application configs:\n\n{compressed_output}\n\n"
"What is the staging database URL?"
),
}
]
resp = _call_claude(messages, [HEADROOM_RETRIEVE_TOOL])
tool_calls = _get_tool_calls(resp)
text = _get_text(resp)
print(f"\n Stop reason: {resp.get('stop_reason')}")
print(f" Tool calls: {tool_calls}")
if text:
print(f" Text: {text[:200]}")
if resp.get("stop_reason") == "tool_use":
print(" RESULT: LLM proactively retrieved (smart)")
else:
print(" RESULT: LLM did NOT retrieve staging config")
def test_specific_user_in_large_list_with_summary(self):
"""Find a specific user in a compressed user list.
Summary mentions user roles. User asks about admins.
"""
visible = [
{"id": i, "name": f"user_{i}", "role": "member", "email": f"user{i}@co.com"}
for i in range(5)
]
all_items = (
visible
+ [
{"id": i, "name": f"user_{i}", "role": "member", "email": f"user{i}@co.com"}
for i in range(5, 95)
]
+ [
{"id": 96, "name": "admin_sarah", "role": "admin", "email": "sarah@co.com"},
{"id": 97, "name": "admin_mike", "role": "admin", "email": "mike@co.com"},
{"id": 98, "name": "superadmin_jane", "role": "superadmin", "email": "jane@co.com"},
]
)
from headroom.transforms.compression_summary import summarize_dropped_items
summary = summarize_dropped_items(all_items, visible)
compressed_output = json.dumps(visible, indent=2)
compressed_output += (
f"\n[{len(all_items) - len(visible)} items compressed to {len(visible)}."
f" Omitted: {summary}."
f' Retrieve: headroom_retrieve(hash="users_hash", query="search")]'
)
messages = [
{
"role": "user",
"content": (
f"Here's our user list:\n\n{compressed_output}\n\n"
"Who are the admin users? I need to contact them."
),
}
]
resp = _call_claude(messages, [HEADROOM_RETRIEVE_TOOL])
tool_calls = _get_tool_calls(resp)
text = _get_text(resp)
print(f"\n Summary: {summary}")
print(f" Stop reason: {resp.get('stop_reason')}")
print(f" Tool calls: {tool_calls}")
if text:
print(f" Text: {text[:200]}")
if resp.get("stop_reason") == "tool_use":
query = tool_calls[0]["input"].get("query", "").lower()
assert "admin" in query
print(f" RESULT: Retrieved admin users (query='{query}') ✓")
else:
print(" RESULT: Did not retrieve admin users")
|