microsoft/TypeAgent
Publicmirrored from https://github.com/microsoft/TypeAgentAvailable
python/ta/test/cmpsearch.py
207lines · modecode
| 1 | # Copyright (c) Microsoft Corporation. |
| 2 | # Licensed under the MIT License. |
| 3 | |
| 4 | import argparse |
| 5 | import asyncio |
| 6 | import builtins |
| 7 | from dataclasses import dataclass |
| 8 | import json |
| 9 | |
| 10 | import numpy as np |
| 11 | import typechat |
| 12 | |
| 13 | from typeagent.aitools import auth |
| 14 | from typeagent.aitools.embeddings import AsyncEmbeddingModel |
| 15 | from typeagent.demo import ui # TODO: Move what we import to a more appropriate place |
| 16 | from typeagent.knowpro.convknowledge import create_typechat_model |
| 17 | from typeagent.knowpro.interfaces import IConversation |
| 18 | from typeagent.knowpro.search_query_schema import SearchQuery |
| 19 | from typeagent.knowpro import searchlang |
| 20 | from typeagent.podcasts.podcast import Podcast |
| 21 | |
| 22 | |
| 23 | @dataclass |
| 24 | class Context: |
| 25 | conversation: IConversation |
| 26 | query_translator: typechat.TypeChatJsonTranslator[SearchQuery] |
| 27 | embedding_model: AsyncEmbeddingModel |
| 28 | options: None |
| 29 | interactive: bool |
| 30 | |
| 31 | |
| 32 | def main(): |
| 33 | # Parse arguments. |
| 34 | |
| 35 | default_qafile = ( |
| 36 | "../../../AISystems-Archive/data/knowpro/test/Episode_53_Answer_results.json" |
| 37 | ) |
| 38 | default_podcast_file = "testdata/Episode_53_AdrianTchaikovsky_index" |
| 39 | |
| 40 | explanation = "a list of objects with 'question' and 'answer' keys" |
| 41 | parser = argparse.ArgumentParser(description="Parse Q/A data file") |
| 42 | parser.add_argument( |
| 43 | "--qafile", |
| 44 | type=str, |
| 45 | default=default_qafile, |
| 46 | help=f"Path to the data file ({explanation})", |
| 47 | ) |
| 48 | parser.add_argument( |
| 49 | "--podcast", |
| 50 | type=str, |
| 51 | default=default_podcast_file, |
| 52 | help="Path to the podcast index files (excluding the '_index.json' suffix)", |
| 53 | ) |
| 54 | parser.add_argument( |
| 55 | "--skip", |
| 56 | type=int, |
| 57 | default=0, |
| 58 | help="Number of initial Q/A pairs to skip (for debugging purposes)", |
| 59 | ) |
| 60 | parser.add_argument( |
| 61 | "--interactive", |
| 62 | "-i", |
| 63 | action="store_true", |
| 64 | default=False, |
| 65 | help="Run in interactive mode, waiting for user input before each question", |
| 66 | ) |
| 67 | args = parser.parse_args() |
| 68 | |
| 69 | # Read evaluation data. |
| 70 | |
| 71 | with open(args.qafile, "r") as file: |
| 72 | data = json.load(file) |
| 73 | assert isinstance(data, list), "Expected a list of Q/A pairs" |
| 74 | assert len(data) > 0, "Expected non-empty Q/A data" |
| 75 | assert all( |
| 76 | isinstance(qa_pair, dict) and "question" in qa_pair and "answer" in qa_pair |
| 77 | for qa_pair in data |
| 78 | ), "Expected each Q/A pair to be a dict with 'question' and 'answer' keys" |
| 79 | |
| 80 | # Read podcast data. |
| 81 | |
| 82 | auth.load_dotenv() |
| 83 | conversation = Podcast.read_from_file(args.podcast) |
| 84 | assert conversation is not None, f"Failed to load podcast from {file!r}" |
| 85 | |
| 86 | # Create translator. |
| 87 | |
| 88 | model = create_typechat_model() |
| 89 | query_translator = ui.create_translator(model, SearchQuery) |
| 90 | |
| 91 | # Create context. |
| 92 | |
| 93 | context = Context( |
| 94 | conversation, |
| 95 | query_translator, |
| 96 | AsyncEmbeddingModel(), |
| 97 | options=None, # TODO: Set options if needed |
| 98 | interactive=args.interactive, |
| 99 | ) |
| 100 | |
| 101 | # Loop over eval data, skipping duplicate questions |
| 102 | # (Those differ in 'cmd' value, which we don't support yet.) |
| 103 | |
| 104 | skip = args.skip |
| 105 | last_q = "" |
| 106 | counter = 0 |
| 107 | for qa_pair in data: |
| 108 | counter += 1 |
| 109 | question = qa_pair.get("question") |
| 110 | answer = qa_pair.get("answer") |
| 111 | if not (question and answer) or question == last_q: |
| 112 | continue |
| 113 | last_q = question |
| 114 | if skip > 0: |
| 115 | skip -= 1 |
| 116 | continue |
| 117 | |
| 118 | # Wait for user input before continuing. |
| 119 | if context.interactive: |
| 120 | try: |
| 121 | input("Press Enter to continue... ") |
| 122 | except (EOFError, KeyboardInterrupt): |
| 123 | print() |
| 124 | break |
| 125 | |
| 126 | # Compare the given answer with the actual answer for the question. |
| 127 | actual_answer, score = asyncio.run(compare(context, qa_pair)) |
| 128 | print("-" * 25, counter, "-" * 25) |
| 129 | if not context.interactive and score < 0.97: |
| 130 | print(f"Question: {question}") |
| 131 | print(f"Expected answer:\n{answer}") |
| 132 | print("-" * 20) |
| 133 | print(f"Actual answer:\n{actual_answer}") |
| 134 | print(f"Score: {score:.3f}") |
| 135 | else: |
| 136 | print(f"Score: {score:.3f} (question: {question})") |
| 137 | |
| 138 | |
| 139 | async def compare( |
| 140 | context: Context, qa_pair: dict[str, str | None] |
| 141 | ) -> tuple[str | None, float]: |
| 142 | the_answer: str | None = None |
| 143 | score = 0.0 |
| 144 | |
| 145 | question = qa_pair.get("question") |
| 146 | answer = qa_pair.get("answer") |
| 147 | cmd = qa_pair.get("cmd") |
| 148 | if not (question and answer): |
| 149 | return None, score |
| 150 | |
| 151 | if not context.interactive: |
| 152 | print = lambda *args, **kwds: None # Disable printing in non-interactive mode |
| 153 | else: |
| 154 | print = builtins.print |
| 155 | |
| 156 | print() |
| 157 | print("=" * 40) |
| 158 | if cmd: |
| 159 | print(f"Command: {cmd}") |
| 160 | print(f"Question: {question}") |
| 161 | print(f"Answer: {answer}") |
| 162 | print("-" * 40) |
| 163 | |
| 164 | result = await searchlang.search_conversation_with_language( |
| 165 | context.conversation, |
| 166 | context.query_translator, |
| 167 | question, |
| 168 | context.options, |
| 169 | ) |
| 170 | print("-" * 40) |
| 171 | if not isinstance(result, typechat.Success): |
| 172 | print("Error:", result.message) |
| 173 | else: |
| 174 | all_answers, combined_answer = await ui.generate_answers( |
| 175 | result.value, context.conversation, question |
| 176 | ) |
| 177 | print("-" * 40) |
| 178 | if combined_answer.type == "NoAnswer": |
| 179 | print("Failure:", combined_answer.whyNoAnswer) |
| 180 | print("All answers:") |
| 181 | if context.interactive: |
| 182 | ui.pretty_print(all_answers) |
| 183 | else: |
| 184 | assert combined_answer.answer is not None, "Expected an answer" |
| 185 | the_answer = combined_answer.answer |
| 186 | score = await equality_score(context, answer, the_answer) |
| 187 | print(the_answer) |
| 188 | print("Correctness score:", score) |
| 189 | print("=" * 40) |
| 190 | |
| 191 | return the_answer, score |
| 192 | |
| 193 | |
| 194 | async def equality_score(context: Context, a: str, b: str) -> float: |
| 195 | a = a.strip() |
| 196 | b = b.strip() |
| 197 | if a == b: |
| 198 | return 1.0 |
| 199 | if a.lower() == b.lower(): |
| 200 | return 0.999 |
| 201 | embeddings = await context.embedding_model.get_embeddings([a, b]) |
| 202 | assert embeddings.shape[0] == 2, "Expected two embeddings" |
| 203 | return np.dot(embeddings[0], embeddings[1]) |
| 204 | |
| 205 | |
| 206 | if __name__ == "__main__": |
| 207 | main() |