-
Notifications
You must be signed in to change notification settings - Fork 0
/
retriever.py
225 lines (200 loc) · 7.75 KB
/
retriever.py
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
from datetime import datetime
from functools import wraps
import json
import logging
import re
import os
import time
import boto3
import botocore
from botocore.config import Config
from boto3.dynamodb.conditions import Key
from sqlalchemy import create_engine, text
retry_config = Config(
retries={"max_attempts": 5, "mode": "standard"},
)
DB_CONNECTION_STRING = os.getenv("DB_CONNECTION_STRING")
logger = logging.getLogger("question-service")
ddb_r = boto3.resource("dynamodb")
bedrock_r = boto3.client("bedrock-runtime", config=retry_config)
prompt_table = ddb_r.Table("session-concierge")
comprehend = boto3.client("comprehend")
engine = create_engine(DB_CONNECTION_STRING)
class DateTimeEncoder(json.JSONEncoder):
def default(self, obj):
if isinstance(obj, datetime):
return obj.isoformat()
return json.JSONEncoder.default(self, obj)
def cache_with_timeout(timeout_sec):
cache = {}
def decorator(func):
@wraps(func)
def wrapped(*args, **kwargs):
now = time.time()
key = args + tuple(sorted(kwargs.items()))
if key not in cache or now - cache[key][1] > timeout_sec:
cache[key] = (func(*args, **kwargs), now)
return cache[key][0]
return wrapped
return decorator
@cache_with_timeout(300)
def fetch_latest_prompt(prompt_table, pk, sk_prefix="PROMPT"):
response = prompt_table.query(
KeyConditionExpression=Key("PK").eq(pk) & Key("SK").begins_with(sk_prefix),
ScanIndexForward=False,
Limit=1,
)
items = response.get("Items", [])
return items[0]["prompt"]
def get_vector_for_query(text_for_vectorization):
return json.loads(
bedrock_r.invoke_model(
modelId="amazon.titan-embed-text-v1",
body=json.dumps({"inputText": text_for_vectorization}),
)["body"].read()
)["embedding"]
def generate_query(query_str):
query = json.loads(
bedrock_r.invoke_model(
modelId="anthropic.claude-v2:1",
body=json.dumps(
{
"temperature": 0.1,
"top_p": 1.0,
"max_tokens_to_sample": 1024 * 10,
"prompt": fetch_latest_prompt(
prompt_table,
"PROMPT#SQL_QUERY_GEN",
).format(
query_str=query_str,
schema=fetch_latest_prompt(
prompt_table,
"PROMPT#TABLE_SCHEMA",
),
),
}
),
)["body"].read()
)["completion"].strip()
if "'[query_vector]'" in query:
query = query.replace(
"'[query_vector]'", f"'{get_vector_for_query(query_str)}'"
)
return re.search(r"<SQL>\s*(.*?)\s*</SQL>", query, re.DOTALL).group(1)
def correct_sql_query(original_question, query, error):
resp = json.loads(
bedrock_r.invoke_model(
modelId="anthropic.claude-v2:1",
body=json.dumps(
{
"temperature": 0.9,
"top_p": 1.0,
"max_tokens_to_sample": 1024 * 10,
"prompt": fetch_latest_prompt(
prompt_table,
"PROMPT#CORRECT_SQL_QUERY",
).format(
original_question=original_question,
original_query=query,
error=error,
schema=fetch_latest_prompt(
prompt_table, "PROMPT#TABLE_SCHEMA", "PROMPT"
),
),
}
),
)["body"].read()
)["completion"].strip()
return re.search(r"<NEW_QUERY>\s*(.*?)\s*</NEW_QUERY>", resp, re.DOTALL).group(1)
def generate_response_stream(original_question, results):
return bedrock_r.invoke_model_with_response_stream(
modelId="anthropic.claude-v2:1",
body=json.dumps(
{
"temperature": 0.9,
"top_p": 1.0,
"max_tokens_to_sample": 1024 * 10,
"prompt": fetch_latest_prompt(
prompt_table, "PROMPT#RESULT_GEN", "PROMPT"
).format(original_question=original_question, results=results),
}
),
).get("body")
def question_stream(
original_question, query=None, retry_on_exception=False, retry_on_no_results=False
):
try:
yield {"status": "Comprehend Classifying Prompt"}
prompt_classes = comprehend.classify_document(
EndpointArn=f"arn:aws:comprehend:{comprehend.meta.region_name}:aws:document-classifier-endpoint/prompt-intent",
Text=original_question,
)["Classes"]
for cls in prompt_classes:
if cls["Name"] == "UNDESIRED_PROMPT":
if cls["Score"] >= 0.9:
yield {
"status": "error",
"output": "Amazon Comprehend determined this was an unsafe prompt. Please try again.",
}
logger.debug(
f"Unsafe prompt. Score: {cls['Score']} {original_question}"
)
return
yield {"status": "Generating Query"}
if query is None:
query = generate_query(original_question)
logger.info(query)
with engine.connect() as conn:
try:
yield {"status": "Executing Query"}
rs = conn.execute(text(query))
yield {"status": "Successful Query"}
except Exception as ex:
if retry_on_exception:
yield {"status": "Error encountered, retrying query"}
logger.debug(f"retrying sql query: {ex}")
new_query = correct_sql_query(original_question, query, ex)
return question_stream(
original_question,
new_query,
retry_on_exception=False,
retry_on_no_results=True,
)
else:
raise
columns = rs.keys()
results = [
{column: row[idx] for idx, column in enumerate(columns)}
for row in rs.fetchall()
]
if not len(results) and retry_on_no_results:
logger.debug("Query had no results, trying again")
yield {"status": "Query had no results, trying again"}
new_query = correct_sql_query(original_question, query, "No results")
return question_stream(
original_question,
new_query,
retry_on_exception=False,
retry_on_no_results=False,
)
logger.debug(results)
if stream := generate_response_stream(original_question, results):
yield {"status": "Generating response"}
complete_response = []
for event in stream:
if chunk := event.get("chunk"):
data = json.loads(chunk.get("bytes").decode())
complete_response.append(data["completion"])
yield data
yield {
"status": "success",
"output": "".join(complete_response).strip(),
}
logger.debug(complete_response)
except botocore.exceptions.EventStreamError:
yield {
"status": "error",
"output": "Bedrock throttled. Please wait and try your question again",
}
except Exception as ex:
yield {"status": "error", "output": f"Error encountered: {ex}"}