|
1 | | -from aimon import Detect |
| 1 | +from aimon import Client, Detect |
2 | 2 | from dataclasses import dataclass |
3 | 3 |
|
4 | 4 | @dataclass |
5 | 5 | class ReactConfig: |
6 | 6 | publish: bool |
| 7 | + async_mode:bool |
7 | 8 | model_name: str |
8 | 9 | max_attempts: int |
9 | 10 | aimon_api_key: str |
10 | 11 | application_name: str |
11 | | - values_returned: list[str] |
12 | 12 | hallucination_threshold: float |
13 | | - aimon_config: dict[str, dict[str, str]] |
| 13 | + framework:str = None |
14 | 14 |
|
15 | | -## ReAct -> Reason and Act |
| 15 | +class React: |
| 16 | + |
| 17 | + ## Initailze the AIMon Client here |
| 18 | + def __init__(self, llm_app, react_configuration, context_extractor): |
| 19 | + |
| 20 | + self.llm_app = llm_app |
| 21 | + self.context_extractor = context_extractor |
| 22 | + self.react_configuration = react_configuration |
| 23 | + self.client = Client(auth_header="Bearer {}".format(self.react_configuration.aimon_api_key)) |
| 24 | + |
| 25 | + def create_payload(self, context, user_query, user_instructions, generated_text): |
| 26 | + |
| 27 | + aimon_payload = { |
| 28 | + 'context':context, |
| 29 | + 'user_query':user_query, |
| 30 | + 'generated_text':generated_text, |
| 31 | + 'instructions':user_instructions, |
| 32 | + } |
| 33 | + |
| 34 | + aimon_payload['publish'] = self.react_configuration.publish |
| 35 | + aimon_payload['async_mode'] = self.react_configuration.async_mode |
| 36 | + aimon_payload['config'] = { 'hallucination': {'detector_name': 'default'}, |
| 37 | + 'instruction_adherence': {'detector_name': 'default'},} |
| 38 | + |
| 39 | + if self.react_configuration.publish: |
| 40 | + aimon_payload['application_name'] = self.react_configuration.application_name |
| 41 | + aimon_payload['model_name'] = self.react_configuration.model_name |
| 42 | + |
| 43 | + return aimon_payload |
16 | 44 |
|
17 | | -def react( llm_app, |
18 | | - user_query, |
19 | | - user_instructions, |
20 | | - context_extractor, |
21 | | - react_configuration, |
22 | | - ): |
| 45 | + |
| 46 | + ## ReAct -> Reason and Act |
| 47 | + def react(self, user_query, user_instructions,): |
| 48 | + |
| 49 | + llm_response = self.llm_app(user_query, user_instructions, reprompted_flag=False) |
| 50 | + |
| 51 | + context = self.context_extractor(user_query, user_instructions, llm_response) |
| 52 | + |
| 53 | + ## Generated text for LLM Response, if the user employs the LlamaIndex framework |
| 54 | + if llm_response.response or self.react_configuration.framework=="llamaindex": |
| 55 | + generated_text = llm_response.response |
| 56 | + else: |
| 57 | + generated_text = llm_response |
23 | 58 |
|
24 | | - detect = Detect(values_returned = react_configuration.values_returned, |
25 | | - api_key = react_configuration.aimon_api_key, |
26 | | - config = react_configuration.aimon_config, |
27 | | - publish = react_configuration.publish, |
28 | | - application_name = react_configuration.application_name, |
29 | | - model_name = react_configuration.model_name, |
30 | | - ) |
31 | | - |
32 | | - llm_response = llm_app(user_query, user_instructions, reprompted_flag=False) |
33 | | - |
34 | | - ## Decorating the context_extractor function with AIMon's "detect" |
35 | | - context_extractor = detect(context_extractor) |
| 59 | + aimon_payload = self.create_payload(context, user_query, user_instructions, generated_text) |
36 | 60 |
|
37 | | - _, _, _, query_result, aimon_response = context_extractor(user_query, user_instructions, llm_response) |
| 61 | + detect_response = self.client.inference.detect(body=[aimon_payload]) |
| 62 | + |
| 63 | + for _ in range(self.react_configuration.max_attempts): |
38 | 64 |
|
39 | | - for _ in range(react_configuration.max_attempts): |
| 65 | + failed_instructions = [] |
| 66 | + ## Loop to check for failed instructions |
| 67 | + for x in detect_response.instruction_adherence['results']: |
| 68 | + if x['adherence'] == False: |
| 69 | + failed_instructions.append(x['instruction']) |
40 | 70 |
|
41 | | - failed_instructions = [] |
42 | | - ## Loop to check for failed instructions |
43 | | - for x in aimon_response.detect_response.instruction_adherence['results']: |
44 | | - if x['adherence'] == False: |
45 | | - failed_instructions.append(x['instruction']) |
| 71 | + hallucination_score = detect_response.hallucination['score'] |
46 | 72 |
|
47 | | - hallucination_score = aimon_response.detect_response.hallucination['score'] |
| 73 | + ## Check whether the hallucination score is greater than the required threshold OR if any of the supplied instructions are not complied with |
| 74 | + if self.react_configuration.hallucination_threshold > 0 and \ |
| 75 | + (hallucination_score > self.react_configuration.hallucination_threshold or len(failed_instructions)>0): |
| 76 | + |
| 77 | + llm_response = self.llm_app(user_query, user_instructions, reprompted_flag=True, hallucination_score=hallucination_score) |
| 78 | + |
| 79 | + context = self.context_extractor(user_query, user_instructions, llm_response) |
| 80 | + |
| 81 | + ## Generated text for LLM Response, if the user employs the LlamaIndex framework |
| 82 | + if llm_response.response or self.react_configuration.framework=="llamaindex": |
| 83 | + generated_text = llm_response.response |
| 84 | + else: |
| 85 | + generated_text = llm_response |
48 | 86 |
|
49 | | - ## Check whether the hallucination score is greater than the required threshold OR if any of the supplied instructions are not complied with |
50 | | - if react_configuration.hallucination_threshold > 0 and \ |
51 | | - (hallucination_score > react_configuration.hallucination_threshold or len(failed_instructions)>0): |
52 | | - |
53 | | - llm_response = llm_app(user_query, user_instructions, reprompted_flag=True, hallucination_score=hallucination_score) |
| 87 | + new_aimon_payload = self.create_payload(context, user_query, user_instructions, generated_text) |
54 | 88 |
|
55 | | - _, _, _, query_result, aimon_response = context_extractor(user_query, user_instructions, llm_response) |
| 89 | + detect_response = self.client.inference.detect(body=[new_aimon_payload]) |
56 | 90 |
|
57 | | - if hallucination_score > react_configuration.hallucination_threshold: |
58 | | - return f"The generated LLM response, even after {react_configuration.max_attempts} attempts of ReAct is still hallucinated. The response: {query_result}" |
| 91 | + if hallucination_score > self.react_configuration.hallucination_threshold: |
| 92 | + return f"The generated LLM response, even after {self.react_configuration.max_attempts} attempts of ReAct is still hallucinated. The response: {generated_text}" |
59 | 93 |
|
60 | | - return query_result |
| 94 | + return generated_text |
0 commit comments