langchain 1.0.0rc2__py3-none-any.whl → 1.0.2__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of langchain might be problematic. Click here for more details.
- langchain/__init__.py +1 -1
- langchain/agents/factory.py +26 -20
- langchain/agents/middleware/__init__.py +12 -0
- langchain/agents/middleware/_execution.py +388 -0
- langchain/agents/middleware/_redaction.py +350 -0
- langchain/agents/middleware/file_search.py +382 -0
- langchain/agents/middleware/pii.py +43 -477
- langchain/agents/middleware/shell_tool.py +718 -0
- langchain/agents/middleware/types.py +7 -5
- langchain/chat_models/base.py +7 -17
- langchain/embeddings/__init__.py +6 -0
- langchain/embeddings/base.py +21 -7
- langchain/tools/tool_node.py +147 -61
- {langchain-1.0.0rc2.dist-info → langchain-1.0.2.dist-info}/METADATA +12 -9
- {langchain-1.0.0rc2.dist-info → langchain-1.0.2.dist-info}/RECORD +17 -13
- {langchain-1.0.0rc2.dist-info → langchain-1.0.2.dist-info}/WHEEL +0 -0
- {langchain-1.0.0rc2.dist-info → langchain-1.0.2.dist-info}/licenses/LICENSE +0 -0
|
@@ -2,15 +2,22 @@
|
|
|
2
2
|
|
|
3
3
|
from __future__ import annotations
|
|
4
4
|
|
|
5
|
-
import hashlib
|
|
6
|
-
import ipaddress
|
|
7
|
-
import re
|
|
8
5
|
from typing import TYPE_CHECKING, Any, Literal
|
|
9
|
-
from urllib.parse import urlparse
|
|
10
6
|
|
|
11
7
|
from langchain_core.messages import AIMessage, AnyMessage, HumanMessage, ToolMessage
|
|
12
|
-
from typing_extensions import TypedDict
|
|
13
8
|
|
|
9
|
+
from langchain.agents.middleware._redaction import (
|
|
10
|
+
PIIDetectionError,
|
|
11
|
+
PIIMatch,
|
|
12
|
+
RedactionRule,
|
|
13
|
+
ResolvedRedactionRule,
|
|
14
|
+
apply_strategy,
|
|
15
|
+
detect_credit_card,
|
|
16
|
+
detect_email,
|
|
17
|
+
detect_ip,
|
|
18
|
+
detect_mac_address,
|
|
19
|
+
detect_url,
|
|
20
|
+
)
|
|
14
21
|
from langchain.agents.middleware.types import AgentMiddleware, AgentState, hook_config
|
|
15
22
|
|
|
16
23
|
if TYPE_CHECKING:
|
|
@@ -19,396 +26,6 @@ if TYPE_CHECKING:
|
|
|
19
26
|
from langgraph.runtime import Runtime
|
|
20
27
|
|
|
21
28
|
|
|
22
|
-
class PIIMatch(TypedDict):
|
|
23
|
-
"""Represents a detected PII match in text."""
|
|
24
|
-
|
|
25
|
-
type: str
|
|
26
|
-
"""The type of PII detected (e.g., 'email', 'ssn', 'credit_card')."""
|
|
27
|
-
value: str
|
|
28
|
-
"""The actual matched text."""
|
|
29
|
-
start: int
|
|
30
|
-
"""Starting position of the match in the text."""
|
|
31
|
-
end: int
|
|
32
|
-
"""Ending position of the match in the text."""
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
class PIIDetectionError(Exception):
|
|
36
|
-
"""Exception raised when PII is detected and strategy is 'block'."""
|
|
37
|
-
|
|
38
|
-
def __init__(self, pii_type: str, matches: list[PIIMatch]) -> None:
|
|
39
|
-
"""Initialize the exception with PII detection information.
|
|
40
|
-
|
|
41
|
-
Args:
|
|
42
|
-
pii_type: The type of PII that was detected.
|
|
43
|
-
matches: List of PII matches found.
|
|
44
|
-
"""
|
|
45
|
-
self.pii_type = pii_type
|
|
46
|
-
self.matches = matches
|
|
47
|
-
count = len(matches)
|
|
48
|
-
msg = f"Detected {count} instance(s) of {pii_type} in message content"
|
|
49
|
-
super().__init__(msg)
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
# ============================================================================
|
|
53
|
-
# PII Detection Functions
|
|
54
|
-
# ============================================================================
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
def _luhn_checksum(card_number: str) -> bool:
|
|
58
|
-
"""Validate credit card number using Luhn algorithm.
|
|
59
|
-
|
|
60
|
-
Args:
|
|
61
|
-
card_number: Credit card number string (digits only).
|
|
62
|
-
|
|
63
|
-
Returns:
|
|
64
|
-
True if the number passes Luhn validation, False otherwise.
|
|
65
|
-
"""
|
|
66
|
-
digits = [int(d) for d in card_number if d.isdigit()]
|
|
67
|
-
|
|
68
|
-
if len(digits) < 13 or len(digits) > 19:
|
|
69
|
-
return False
|
|
70
|
-
|
|
71
|
-
checksum = 0
|
|
72
|
-
for i, digit in enumerate(reversed(digits)):
|
|
73
|
-
d = digit
|
|
74
|
-
if i % 2 == 1:
|
|
75
|
-
d *= 2
|
|
76
|
-
if d > 9:
|
|
77
|
-
d -= 9
|
|
78
|
-
checksum += d
|
|
79
|
-
|
|
80
|
-
return checksum % 10 == 0
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
def detect_email(content: str) -> list[PIIMatch]:
|
|
84
|
-
"""Detect email addresses in content.
|
|
85
|
-
|
|
86
|
-
Args:
|
|
87
|
-
content: Text content to scan.
|
|
88
|
-
|
|
89
|
-
Returns:
|
|
90
|
-
List of detected email matches.
|
|
91
|
-
"""
|
|
92
|
-
pattern = r"\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b"
|
|
93
|
-
return [
|
|
94
|
-
PIIMatch(
|
|
95
|
-
type="email",
|
|
96
|
-
value=match.group(),
|
|
97
|
-
start=match.start(),
|
|
98
|
-
end=match.end(),
|
|
99
|
-
)
|
|
100
|
-
for match in re.finditer(pattern, content)
|
|
101
|
-
]
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
def detect_credit_card(content: str) -> list[PIIMatch]:
|
|
105
|
-
"""Detect credit card numbers in content using Luhn validation.
|
|
106
|
-
|
|
107
|
-
Detects cards in formats like:
|
|
108
|
-
- 1234567890123456
|
|
109
|
-
- 1234 5678 9012 3456
|
|
110
|
-
- 1234-5678-9012-3456
|
|
111
|
-
|
|
112
|
-
Args:
|
|
113
|
-
content: Text content to scan.
|
|
114
|
-
|
|
115
|
-
Returns:
|
|
116
|
-
List of detected credit card matches.
|
|
117
|
-
"""
|
|
118
|
-
# Match various credit card formats
|
|
119
|
-
pattern = r"\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b"
|
|
120
|
-
matches = []
|
|
121
|
-
|
|
122
|
-
for match in re.finditer(pattern, content):
|
|
123
|
-
card_number = match.group()
|
|
124
|
-
# Validate with Luhn algorithm
|
|
125
|
-
if _luhn_checksum(card_number):
|
|
126
|
-
matches.append(
|
|
127
|
-
PIIMatch(
|
|
128
|
-
type="credit_card",
|
|
129
|
-
value=card_number,
|
|
130
|
-
start=match.start(),
|
|
131
|
-
end=match.end(),
|
|
132
|
-
)
|
|
133
|
-
)
|
|
134
|
-
|
|
135
|
-
return matches
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
def detect_ip(content: str) -> list[PIIMatch]:
|
|
139
|
-
"""Detect IP addresses in content using stdlib validation.
|
|
140
|
-
|
|
141
|
-
Validates both IPv4 and IPv6 addresses.
|
|
142
|
-
|
|
143
|
-
Args:
|
|
144
|
-
content: Text content to scan.
|
|
145
|
-
|
|
146
|
-
Returns:
|
|
147
|
-
List of detected IP address matches.
|
|
148
|
-
"""
|
|
149
|
-
matches = []
|
|
150
|
-
|
|
151
|
-
# IPv4 pattern
|
|
152
|
-
ipv4_pattern = r"\b(?:[0-9]{1,3}\.){3}[0-9]{1,3}\b"
|
|
153
|
-
|
|
154
|
-
for match in re.finditer(ipv4_pattern, content):
|
|
155
|
-
ip_str = match.group()
|
|
156
|
-
try:
|
|
157
|
-
# Validate with stdlib
|
|
158
|
-
ipaddress.ip_address(ip_str)
|
|
159
|
-
matches.append(
|
|
160
|
-
PIIMatch(
|
|
161
|
-
type="ip",
|
|
162
|
-
value=ip_str,
|
|
163
|
-
start=match.start(),
|
|
164
|
-
end=match.end(),
|
|
165
|
-
)
|
|
166
|
-
)
|
|
167
|
-
except ValueError:
|
|
168
|
-
# Not a valid IP address
|
|
169
|
-
pass
|
|
170
|
-
|
|
171
|
-
return matches
|
|
172
|
-
|
|
173
|
-
|
|
174
|
-
def detect_mac_address(content: str) -> list[PIIMatch]:
|
|
175
|
-
"""Detect MAC addresses in content.
|
|
176
|
-
|
|
177
|
-
Detects formats like:
|
|
178
|
-
- 00:1A:2B:3C:4D:5E
|
|
179
|
-
- 00-1A-2B-3C-4D-5E
|
|
180
|
-
|
|
181
|
-
Args:
|
|
182
|
-
content: Text content to scan.
|
|
183
|
-
|
|
184
|
-
Returns:
|
|
185
|
-
List of detected MAC address matches.
|
|
186
|
-
"""
|
|
187
|
-
pattern = r"\b([0-9A-Fa-f]{2}[:-]){5}[0-9A-Fa-f]{2}\b"
|
|
188
|
-
return [
|
|
189
|
-
PIIMatch(
|
|
190
|
-
type="mac_address",
|
|
191
|
-
value=match.group(),
|
|
192
|
-
start=match.start(),
|
|
193
|
-
end=match.end(),
|
|
194
|
-
)
|
|
195
|
-
for match in re.finditer(pattern, content)
|
|
196
|
-
]
|
|
197
|
-
|
|
198
|
-
|
|
199
|
-
def detect_url(content: str) -> list[PIIMatch]:
|
|
200
|
-
"""Detect URLs in content using regex and stdlib validation.
|
|
201
|
-
|
|
202
|
-
Detects:
|
|
203
|
-
- http://example.com
|
|
204
|
-
- https://example.com/path
|
|
205
|
-
- www.example.com
|
|
206
|
-
- example.com/path
|
|
207
|
-
|
|
208
|
-
Args:
|
|
209
|
-
content: Text content to scan.
|
|
210
|
-
|
|
211
|
-
Returns:
|
|
212
|
-
List of detected URL matches.
|
|
213
|
-
"""
|
|
214
|
-
matches = []
|
|
215
|
-
|
|
216
|
-
# Pattern 1: URLs with scheme (http:// or https://)
|
|
217
|
-
scheme_pattern = r"https?://[^\s<>\"{}|\\^`\[\]]+"
|
|
218
|
-
|
|
219
|
-
for match in re.finditer(scheme_pattern, content):
|
|
220
|
-
url = match.group()
|
|
221
|
-
try:
|
|
222
|
-
result = urlparse(url)
|
|
223
|
-
if result.scheme in ("http", "https") and result.netloc:
|
|
224
|
-
matches.append(
|
|
225
|
-
PIIMatch(
|
|
226
|
-
type="url",
|
|
227
|
-
value=url,
|
|
228
|
-
start=match.start(),
|
|
229
|
-
end=match.end(),
|
|
230
|
-
)
|
|
231
|
-
)
|
|
232
|
-
except Exception: # noqa: S110, BLE001
|
|
233
|
-
# Invalid URL, skip
|
|
234
|
-
pass
|
|
235
|
-
|
|
236
|
-
# Pattern 2: URLs without scheme (www.example.com or example.com/path)
|
|
237
|
-
# More conservative to avoid false positives
|
|
238
|
-
bare_pattern = r"\b(?:www\.)?[a-zA-Z0-9](?:[a-zA-Z0-9-]{0,61}[a-zA-Z0-9])?(?:\.[a-zA-Z0-9](?:[a-zA-Z0-9-]{0,61}[a-zA-Z0-9])?)+(?:/[^\s]*)?" # noqa: E501
|
|
239
|
-
|
|
240
|
-
for match in re.finditer(bare_pattern, content):
|
|
241
|
-
# Skip if already matched with scheme
|
|
242
|
-
if any(
|
|
243
|
-
m["start"] <= match.start() < m["end"] or m["start"] < match.end() <= m["end"]
|
|
244
|
-
for m in matches
|
|
245
|
-
):
|
|
246
|
-
continue
|
|
247
|
-
|
|
248
|
-
url = match.group()
|
|
249
|
-
# Only accept if it has a path or starts with www
|
|
250
|
-
# This reduces false positives like "example.com" in prose
|
|
251
|
-
if "/" in url or url.startswith("www."):
|
|
252
|
-
try:
|
|
253
|
-
# Add scheme for validation (required for urlparse to work correctly)
|
|
254
|
-
test_url = f"http://{url}"
|
|
255
|
-
result = urlparse(test_url)
|
|
256
|
-
if result.netloc and "." in result.netloc:
|
|
257
|
-
matches.append(
|
|
258
|
-
PIIMatch(
|
|
259
|
-
type="url",
|
|
260
|
-
value=url,
|
|
261
|
-
start=match.start(),
|
|
262
|
-
end=match.end(),
|
|
263
|
-
)
|
|
264
|
-
)
|
|
265
|
-
except Exception: # noqa: S110, BLE001
|
|
266
|
-
# Invalid URL, skip
|
|
267
|
-
pass
|
|
268
|
-
|
|
269
|
-
return matches
|
|
270
|
-
|
|
271
|
-
|
|
272
|
-
# Built-in detector registry
|
|
273
|
-
_BUILTIN_DETECTORS: dict[str, Callable[[str], list[PIIMatch]]] = {
|
|
274
|
-
"email": detect_email,
|
|
275
|
-
"credit_card": detect_credit_card,
|
|
276
|
-
"ip": detect_ip,
|
|
277
|
-
"mac_address": detect_mac_address,
|
|
278
|
-
"url": detect_url,
|
|
279
|
-
}
|
|
280
|
-
|
|
281
|
-
|
|
282
|
-
# ============================================================================
|
|
283
|
-
# Strategy Implementations
|
|
284
|
-
# ============================================================================
|
|
285
|
-
|
|
286
|
-
|
|
287
|
-
def _apply_redact_strategy(content: str, matches: list[PIIMatch]) -> str:
|
|
288
|
-
"""Replace PII with [REDACTED_TYPE] placeholders.
|
|
289
|
-
|
|
290
|
-
Args:
|
|
291
|
-
content: Original content.
|
|
292
|
-
matches: List of PII matches to redact.
|
|
293
|
-
|
|
294
|
-
Returns:
|
|
295
|
-
Content with PII redacted.
|
|
296
|
-
"""
|
|
297
|
-
if not matches:
|
|
298
|
-
return content
|
|
299
|
-
|
|
300
|
-
# Sort matches by start position in reverse to avoid offset issues
|
|
301
|
-
sorted_matches = sorted(matches, key=lambda m: m["start"], reverse=True)
|
|
302
|
-
|
|
303
|
-
result = content
|
|
304
|
-
for match in sorted_matches:
|
|
305
|
-
replacement = f"[REDACTED_{match['type'].upper()}]"
|
|
306
|
-
result = result[: match["start"]] + replacement + result[match["end"] :]
|
|
307
|
-
|
|
308
|
-
return result
|
|
309
|
-
|
|
310
|
-
|
|
311
|
-
def _apply_mask_strategy(content: str, matches: list[PIIMatch]) -> str:
|
|
312
|
-
"""Partially mask PII, showing only last few characters.
|
|
313
|
-
|
|
314
|
-
Args:
|
|
315
|
-
content: Original content.
|
|
316
|
-
matches: List of PII matches to mask.
|
|
317
|
-
|
|
318
|
-
Returns:
|
|
319
|
-
Content with PII masked.
|
|
320
|
-
"""
|
|
321
|
-
if not matches:
|
|
322
|
-
return content
|
|
323
|
-
|
|
324
|
-
# Sort matches by start position in reverse
|
|
325
|
-
sorted_matches = sorted(matches, key=lambda m: m["start"], reverse=True)
|
|
326
|
-
|
|
327
|
-
result = content
|
|
328
|
-
for match in sorted_matches:
|
|
329
|
-
value = match["value"]
|
|
330
|
-
pii_type = match["type"]
|
|
331
|
-
|
|
332
|
-
# Different masking strategies by type
|
|
333
|
-
if pii_type == "email":
|
|
334
|
-
# Show only domain: user@****.com
|
|
335
|
-
parts = value.split("@")
|
|
336
|
-
if len(parts) == 2:
|
|
337
|
-
domain_parts = parts[1].split(".")
|
|
338
|
-
if len(domain_parts) >= 2:
|
|
339
|
-
masked = f"{parts[0]}@****.{domain_parts[-1]}"
|
|
340
|
-
else:
|
|
341
|
-
masked = f"{parts[0]}@****"
|
|
342
|
-
else:
|
|
343
|
-
masked = "****"
|
|
344
|
-
|
|
345
|
-
elif pii_type == "credit_card":
|
|
346
|
-
# Show last 4: ****-****-****-1234
|
|
347
|
-
digits_only = "".join(c for c in value if c.isdigit())
|
|
348
|
-
separator = "-" if "-" in value else " " if " " in value else ""
|
|
349
|
-
if separator:
|
|
350
|
-
masked = f"****{separator}****{separator}****{separator}{digits_only[-4:]}"
|
|
351
|
-
else:
|
|
352
|
-
masked = f"************{digits_only[-4:]}"
|
|
353
|
-
|
|
354
|
-
elif pii_type == "ip":
|
|
355
|
-
# Show last octet: *.*.*. 123
|
|
356
|
-
parts = value.split(".")
|
|
357
|
-
masked = f"*.*.*.{parts[-1]}" if len(parts) == 4 else "****"
|
|
358
|
-
|
|
359
|
-
elif pii_type == "mac_address":
|
|
360
|
-
# Show last byte: **:**:**:**:**:5E
|
|
361
|
-
separator = ":" if ":" in value else "-"
|
|
362
|
-
masked = (
|
|
363
|
-
f"**{separator}**{separator}**{separator}**{separator}**{separator}{value[-2:]}"
|
|
364
|
-
)
|
|
365
|
-
|
|
366
|
-
elif pii_type == "url":
|
|
367
|
-
# Mask everything: [MASKED_URL]
|
|
368
|
-
masked = "[MASKED_URL]"
|
|
369
|
-
|
|
370
|
-
else:
|
|
371
|
-
# Default: show last 4 chars
|
|
372
|
-
masked = f"****{value[-4:]}" if len(value) > 4 else "****"
|
|
373
|
-
|
|
374
|
-
result = result[: match["start"]] + masked + result[match["end"] :]
|
|
375
|
-
|
|
376
|
-
return result
|
|
377
|
-
|
|
378
|
-
|
|
379
|
-
def _apply_hash_strategy(content: str, matches: list[PIIMatch]) -> str:
|
|
380
|
-
"""Replace PII with deterministic hash including type information.
|
|
381
|
-
|
|
382
|
-
Args:
|
|
383
|
-
content: Original content.
|
|
384
|
-
matches: List of PII matches to hash.
|
|
385
|
-
|
|
386
|
-
Returns:
|
|
387
|
-
Content with PII replaced by hashes in format <type_hash:digest>.
|
|
388
|
-
"""
|
|
389
|
-
if not matches:
|
|
390
|
-
return content
|
|
391
|
-
|
|
392
|
-
# Sort matches by start position in reverse
|
|
393
|
-
sorted_matches = sorted(matches, key=lambda m: m["start"], reverse=True)
|
|
394
|
-
|
|
395
|
-
result = content
|
|
396
|
-
for match in sorted_matches:
|
|
397
|
-
value = match["value"]
|
|
398
|
-
pii_type = match["type"]
|
|
399
|
-
# Create deterministic hash
|
|
400
|
-
hash_digest = hashlib.sha256(value.encode()).hexdigest()[:8]
|
|
401
|
-
replacement = f"<{pii_type}_hash:{hash_digest}>"
|
|
402
|
-
result = result[: match["start"]] + replacement + result[match["end"] :]
|
|
403
|
-
|
|
404
|
-
return result
|
|
405
|
-
|
|
406
|
-
|
|
407
|
-
# ============================================================================
|
|
408
|
-
# PIIMiddleware
|
|
409
|
-
# ============================================================================
|
|
410
|
-
|
|
411
|
-
|
|
412
29
|
class PIIMiddleware(AgentMiddleware):
|
|
413
30
|
"""Detect and handle Personally Identifiable Information (PII) in agent conversations.
|
|
414
31
|
|
|
@@ -510,50 +127,34 @@ class PIIMiddleware(AgentMiddleware):
|
|
|
510
127
|
"""
|
|
511
128
|
super().__init__()
|
|
512
129
|
|
|
513
|
-
self.pii_type = pii_type
|
|
514
|
-
self.strategy = strategy
|
|
515
130
|
self.apply_to_input = apply_to_input
|
|
516
131
|
self.apply_to_output = apply_to_output
|
|
517
132
|
self.apply_to_tool_results = apply_to_tool_results
|
|
518
133
|
|
|
519
|
-
|
|
520
|
-
|
|
521
|
-
|
|
522
|
-
|
|
523
|
-
|
|
524
|
-
|
|
525
|
-
|
|
526
|
-
|
|
527
|
-
)
|
|
528
|
-
raise ValueError(msg)
|
|
529
|
-
self.detector = _BUILTIN_DETECTORS[pii_type]
|
|
530
|
-
elif isinstance(detector, str):
|
|
531
|
-
# Custom regex pattern
|
|
532
|
-
pattern = detector
|
|
533
|
-
|
|
534
|
-
def regex_detector(content: str) -> list[PIIMatch]:
|
|
535
|
-
return [
|
|
536
|
-
PIIMatch(
|
|
537
|
-
type=pii_type,
|
|
538
|
-
value=match.group(),
|
|
539
|
-
start=match.start(),
|
|
540
|
-
end=match.end(),
|
|
541
|
-
)
|
|
542
|
-
for match in re.finditer(pattern, content)
|
|
543
|
-
]
|
|
544
|
-
|
|
545
|
-
self.detector = regex_detector
|
|
546
|
-
else:
|
|
547
|
-
# Custom callable detector
|
|
548
|
-
self.detector = detector
|
|
134
|
+
self._resolved_rule: ResolvedRedactionRule = RedactionRule(
|
|
135
|
+
pii_type=pii_type,
|
|
136
|
+
strategy=strategy,
|
|
137
|
+
detector=detector,
|
|
138
|
+
).resolve()
|
|
139
|
+
self.pii_type = self._resolved_rule.pii_type
|
|
140
|
+
self.strategy = self._resolved_rule.strategy
|
|
141
|
+
self.detector = self._resolved_rule.detector
|
|
549
142
|
|
|
550
143
|
@property
|
|
551
144
|
def name(self) -> str:
|
|
552
145
|
"""Name of the middleware."""
|
|
553
146
|
return f"{self.__class__.__name__}[{self.pii_type}]"
|
|
554
147
|
|
|
148
|
+
def _process_content(self, content: str) -> tuple[str, list[PIIMatch]]:
|
|
149
|
+
"""Apply the configured redaction rule to the provided content."""
|
|
150
|
+
matches = self.detector(content)
|
|
151
|
+
if not matches:
|
|
152
|
+
return content, []
|
|
153
|
+
sanitized = apply_strategy(content, matches, self.strategy)
|
|
154
|
+
return sanitized, matches
|
|
155
|
+
|
|
555
156
|
@hook_config(can_jump_to=["end"])
|
|
556
|
-
def before_model(
|
|
157
|
+
def before_model(
|
|
557
158
|
self,
|
|
558
159
|
state: AgentState,
|
|
559
160
|
runtime: Runtime, # noqa: ARG002
|
|
@@ -594,25 +195,9 @@ class PIIMiddleware(AgentMiddleware):
|
|
|
594
195
|
if last_user_idx is not None and last_user_msg and last_user_msg.content:
|
|
595
196
|
# Detect PII in message content
|
|
596
197
|
content = str(last_user_msg.content)
|
|
597
|
-
matches = self.
|
|
198
|
+
new_content, matches = self._process_content(content)
|
|
598
199
|
|
|
599
200
|
if matches:
|
|
600
|
-
# Apply strategy
|
|
601
|
-
if self.strategy == "block":
|
|
602
|
-
raise PIIDetectionError(self.pii_type, matches)
|
|
603
|
-
|
|
604
|
-
if self.strategy == "redact":
|
|
605
|
-
new_content = _apply_redact_strategy(content, matches)
|
|
606
|
-
elif self.strategy == "mask":
|
|
607
|
-
new_content = _apply_mask_strategy(content, matches)
|
|
608
|
-
elif self.strategy == "hash":
|
|
609
|
-
new_content = _apply_hash_strategy(content, matches)
|
|
610
|
-
else:
|
|
611
|
-
# Should not reach here due to type hints
|
|
612
|
-
msg = f"Unknown strategy: {self.strategy}"
|
|
613
|
-
raise ValueError(msg)
|
|
614
|
-
|
|
615
|
-
# Create updated message
|
|
616
201
|
updated_message: AnyMessage = HumanMessage(
|
|
617
202
|
content=new_content,
|
|
618
203
|
id=last_user_msg.id,
|
|
@@ -641,26 +226,11 @@ class PIIMiddleware(AgentMiddleware):
|
|
|
641
226
|
continue
|
|
642
227
|
|
|
643
228
|
content = str(tool_msg.content)
|
|
644
|
-
matches = self.
|
|
229
|
+
new_content, matches = self._process_content(content)
|
|
645
230
|
|
|
646
231
|
if not matches:
|
|
647
232
|
continue
|
|
648
233
|
|
|
649
|
-
# Apply strategy
|
|
650
|
-
if self.strategy == "block":
|
|
651
|
-
raise PIIDetectionError(self.pii_type, matches)
|
|
652
|
-
|
|
653
|
-
if self.strategy == "redact":
|
|
654
|
-
new_content = _apply_redact_strategy(content, matches)
|
|
655
|
-
elif self.strategy == "mask":
|
|
656
|
-
new_content = _apply_mask_strategy(content, matches)
|
|
657
|
-
elif self.strategy == "hash":
|
|
658
|
-
new_content = _apply_hash_strategy(content, matches)
|
|
659
|
-
else:
|
|
660
|
-
# Should not reach here due to type hints
|
|
661
|
-
msg = f"Unknown strategy: {self.strategy}"
|
|
662
|
-
raise ValueError(msg)
|
|
663
|
-
|
|
664
234
|
# Create updated tool message
|
|
665
235
|
updated_message = ToolMessage(
|
|
666
236
|
content=new_content,
|
|
@@ -716,26 +286,11 @@ class PIIMiddleware(AgentMiddleware):
|
|
|
716
286
|
|
|
717
287
|
# Detect PII in message content
|
|
718
288
|
content = str(last_ai_msg.content)
|
|
719
|
-
matches = self.
|
|
289
|
+
new_content, matches = self._process_content(content)
|
|
720
290
|
|
|
721
291
|
if not matches:
|
|
722
292
|
return None
|
|
723
293
|
|
|
724
|
-
# Apply strategy
|
|
725
|
-
if self.strategy == "block":
|
|
726
|
-
raise PIIDetectionError(self.pii_type, matches)
|
|
727
|
-
|
|
728
|
-
if self.strategy == "redact":
|
|
729
|
-
new_content = _apply_redact_strategy(content, matches)
|
|
730
|
-
elif self.strategy == "mask":
|
|
731
|
-
new_content = _apply_mask_strategy(content, matches)
|
|
732
|
-
elif self.strategy == "hash":
|
|
733
|
-
new_content = _apply_hash_strategy(content, matches)
|
|
734
|
-
else:
|
|
735
|
-
# Should not reach here due to type hints
|
|
736
|
-
msg = f"Unknown strategy: {self.strategy}"
|
|
737
|
-
raise ValueError(msg)
|
|
738
|
-
|
|
739
294
|
# Create updated message
|
|
740
295
|
updated_message = AIMessage(
|
|
741
296
|
content=new_content,
|
|
@@ -749,3 +304,14 @@ class PIIMiddleware(AgentMiddleware):
|
|
|
749
304
|
new_messages[last_ai_idx] = updated_message
|
|
750
305
|
|
|
751
306
|
return {"messages": new_messages}
|
|
307
|
+
|
|
308
|
+
|
|
309
|
+
__all__ = [
|
|
310
|
+
"PIIDetectionError",
|
|
311
|
+
"PIIMiddleware",
|
|
312
|
+
"detect_credit_card",
|
|
313
|
+
"detect_email",
|
|
314
|
+
"detect_ip",
|
|
315
|
+
"detect_mac_address",
|
|
316
|
+
"detect_url",
|
|
317
|
+
]
|