microsoft/TypeAgent
Publicmirrored from https://github.com/microsoft/TypeAgentAvailable
python/ta/tools/utool.py
798lines · modecode
| 1 | # Copyright (c) Microsoft Corporation. |
| 2 | # Licensed under the MIT License. |
| 3 | |
| 4 | __version__ = "0.2" |
| 5 | |
| 6 | ### Imports ### |
| 7 | |
| 8 | import argparse |
| 9 | import asyncio |
| 10 | from collections.abc import Mapping |
| 11 | from dataclasses import dataclass |
| 12 | import difflib |
| 13 | import json |
| 14 | import re |
| 15 | import shutil |
| 16 | import sys |
| 17 | import typing |
| 18 | |
| 19 | from colorama import init as colorama_init, Fore |
| 20 | import numpy as np |
| 21 | |
| 22 | try: |
| 23 | import readline |
| 24 | except ImportError: |
| 25 | readline = None |
| 26 | import typechat |
| 27 | |
| 28 | from typeagent.aitools import embeddings |
| 29 | from typeagent.aitools import utils |
| 30 | |
| 31 | from typeagent.knowpro.interfaces import ( |
| 32 | IConversation, |
| 33 | IMessage, |
| 34 | ITermToSemanticRefIndex, |
| 35 | ScoredMessageOrdinal, |
| 36 | ScoredSemanticRefOrdinal, |
| 37 | SemanticRef, |
| 38 | Topic, |
| 39 | ) |
| 40 | from typeagent.knowpro import answers, answer_response_schema |
| 41 | from typeagent.knowpro import convknowledge |
| 42 | from typeagent.knowpro import importing |
| 43 | from typeagent.knowpro import kplib |
| 44 | from typeagent.knowpro import query |
| 45 | from typeagent.knowpro import search, search_query_schema, searchlang |
| 46 | from typeagent.knowpro import serialization |
| 47 | |
| 48 | from typeagent.podcasts import podcast |
| 49 | |
| 50 | |
| 51 | ### Classes ### |
| 52 | |
| 53 | |
| 54 | class QuestionAnswerData(typing.TypedDict): |
| 55 | question: str |
| 56 | answer: str |
| 57 | hasNoAnswer: bool |
| 58 | cmd: str |
| 59 | |
| 60 | |
| 61 | class RawSearchResultData(typing.TypedDict): |
| 62 | messageMatches: list[int] |
| 63 | entityMatches: list[int] |
| 64 | topicMatches: list[int] |
| 65 | actionMatches: list[int] |
| 66 | |
| 67 | |
| 68 | class SearchResultData(typing.TypedDict): |
| 69 | searchText: str |
| 70 | searchQueryExpr: dict[str, typing.Any] # Serialized search_query_schema.SearchQuery |
| 71 | compiledQueryExpr: list[dict[str, typing.Any]] # list[search.SearchQueryExpr] |
| 72 | results: list[RawSearchResultData] |
| 73 | |
| 74 | |
| 75 | @dataclass |
| 76 | class ProcessingContext: |
| 77 | query_context: query.QueryEvalContext |
| 78 | ar_list: list[QuestionAnswerData] |
| 79 | sr_list: list[SearchResultData] |
| 80 | ar_index: dict[str, QuestionAnswerData] |
| 81 | sr_index: dict[str, SearchResultData] |
| 82 | debug1: typing.Literal["none", "diff", "full", "skip"] |
| 83 | debug2: typing.Literal["none", "diff", "full", "skip"] |
| 84 | debug3: typing.Literal["none", "diff", "full", "nice"] |
| 85 | debug4: typing.Literal["none", "diff", "full", "nice"] |
| 86 | embedding_model: embeddings.AsyncEmbeddingModel |
| 87 | query_translator: typechat.TypeChatJsonTranslator[search_query_schema.SearchQuery] |
| 88 | answer_translator: typechat.TypeChatJsonTranslator[ |
| 89 | answer_response_schema.AnswerResponse |
| 90 | ] |
| 91 | lang_search_options: searchlang.LanguageSearchOptions |
| 92 | answer_context_options: answers.AnswerContextOptions |
| 93 | |
| 94 | def __repr__(self) -> str: |
| 95 | parts = [] |
| 96 | parts.append(f"ar_list={len(self.ar_list)}") |
| 97 | parts.append(f"sr_list={len(self.sr_list)}") |
| 98 | parts.append(f"ar_index={len(self.ar_index)}") |
| 99 | parts.append(f"sr_index={len(self.sr_index)}") |
| 100 | parts.append(f"debug1={self.debug1}") |
| 101 | parts.append(f"debug2={self.debug2}") |
| 102 | parts.append(f"debug3={self.debug3}") |
| 103 | parts.append(f"debug4={self.debug4}") |
| 104 | parts.append(f"lang_search_options={self.lang_search_options}") |
| 105 | parts.append(f"answer_context_options={self.answer_context_options}") |
| 106 | return f"Context({', '.join(parts)})" |
| 107 | |
| 108 | |
| 109 | ### Main logic ### |
| 110 | |
| 111 | |
| 112 | def main(): |
| 113 | utils.load_dotenv() |
| 114 | colorama_init(autoreset=True) |
| 115 | |
| 116 | parser = make_arg_parser("TypeAgent Query Tool") |
| 117 | args = parser.parse_args() |
| 118 | fill_in_debug_defaults(parser, args) |
| 119 | settings = importing.ConversationSettings() |
| 120 | query_context = load_podcast_index(args.podcast, settings) |
| 121 | ar_list, ar_index = load_index_file(args.qafile, "question", QuestionAnswerData) |
| 122 | sr_list, sr_index = load_index_file(args.srfile, "searchText", SearchResultData) |
| 123 | |
| 124 | model = convknowledge.create_typechat_model() |
| 125 | query_translator = utils.create_translator(model, search_query_schema.SearchQuery) |
| 126 | if args.alt_schema: |
| 127 | print(f"Substituting alt schema from {args.alt_schema}") |
| 128 | with open(args.alt_schema) as f: |
| 129 | query_translator.schema_str = f.read() |
| 130 | if args.show_schema: |
| 131 | print(Fore.YELLOW + query_translator.schema_str.rstrip() + Fore.RESET) |
| 132 | |
| 133 | answer_translator = utils.create_translator( |
| 134 | model, answer_response_schema.AnswerResponse |
| 135 | ) |
| 136 | |
| 137 | context = ProcessingContext( |
| 138 | query_context, |
| 139 | ar_list, |
| 140 | sr_list, |
| 141 | ar_index, |
| 142 | sr_index, |
| 143 | args.debug1, |
| 144 | args.debug2, |
| 145 | args.debug3, |
| 146 | args.debug4, |
| 147 | settings.embedding_model, |
| 148 | query_translator, |
| 149 | answer_translator, |
| 150 | searchlang.LanguageSearchOptions( |
| 151 | compile_options=searchlang.LanguageQueryCompileOptions( |
| 152 | exact_scope=False, verb_scope=True, term_filter=None, apply_scope=True |
| 153 | ), |
| 154 | exact_match=False, |
| 155 | max_message_matches=25, |
| 156 | ), |
| 157 | answers.AnswerContextOptions( |
| 158 | entities_top_k=50, topics_top_k=50, messages_top_k=None, chunking=None |
| 159 | ), |
| 160 | ) |
| 161 | |
| 162 | utils.pretty_print(context, Fore.BLUE, Fore.RESET) |
| 163 | |
| 164 | if args.batch: |
| 165 | print( |
| 166 | Fore.YELLOW |
| 167 | + f"Running in batch mode [{args.offset}:{args.offset + args.limit if args.limit else ''}]." |
| 168 | + Fore.RESET |
| 169 | ) |
| 170 | asyncio.run(batch_loop(context, args.offset, args.limit)) |
| 171 | else: |
| 172 | print(Fore.YELLOW + "Running in interactive mode." + Fore.RESET) |
| 173 | interactive_loop(context) |
| 174 | |
| 175 | |
| 176 | async def batch_loop(context: ProcessingContext, offset: int, limit: int) -> None: |
| 177 | if limit == 0: |
| 178 | limit = len(context.ar_list) - offset |
| 179 | sublist = context.ar_list[offset : offset + limit] |
| 180 | all_scores = [] |
| 181 | for counter, qadata in enumerate(sublist, offset + 1): |
| 182 | question = qadata["question"] |
| 183 | print("-" * 20, counter, question, "-" * 20) |
| 184 | score = await process_query(context, question) |
| 185 | if score is not None: |
| 186 | all_scores.append((score, counter)) |
| 187 | if not all_scores: |
| 188 | return |
| 189 | print("=" * 50) |
| 190 | all_scores.sort(reverse=True) |
| 191 | good_scores = [(score, counter) for score, counter in all_scores if score >= 0.97] |
| 192 | bad_scores = [(score, counter) for score, counter in all_scores if score < 0.97] |
| 193 | for label, pairs in [("Good", good_scores), ("Bad", bad_scores)]: |
| 194 | print(f"{label} scores ({len(pairs)}):") |
| 195 | for i in range(0, len(pairs), 10): |
| 196 | print( |
| 197 | ", ".join( |
| 198 | f"{score:.3f}({counter})" for score, counter in pairs[i : i + 10] |
| 199 | ) |
| 200 | ) |
| 201 | |
| 202 | |
| 203 | def interactive_loop(context: ProcessingContext) -> None: |
| 204 | if not sys.stdin.isatty(): |
| 205 | for line in sys.stdin: |
| 206 | line = line.strip() |
| 207 | if not line: |
| 208 | continue |
| 209 | asyncio.run(process_query(context, line)) |
| 210 | return |
| 211 | |
| 212 | print(f"TypeAgent demo UI {__version__} (type 'q' to exit)") |
| 213 | if readline: |
| 214 | try: |
| 215 | readline.read_history_file(".ui_history") |
| 216 | except FileNotFoundError: |
| 217 | pass # Ignore if history file does not exist. |
| 218 | |
| 219 | try: |
| 220 | while True: |
| 221 | try: |
| 222 | line = input("TypeAgent> ").strip() |
| 223 | except EOFError: |
| 224 | print() |
| 225 | break |
| 226 | if not line: |
| 227 | continue |
| 228 | if line.lower() in ("exit", "quit", "q"): |
| 229 | if readline: |
| 230 | readline.remove_history_item( |
| 231 | readline.get_current_history_length() - 1 |
| 232 | ) |
| 233 | break |
| 234 | prsep() |
| 235 | asyncio.run(process_query(context, line)) |
| 236 | |
| 237 | finally: |
| 238 | if readline: |
| 239 | readline.write_history_file(".ui_history") |
| 240 | |
| 241 | |
| 242 | ### Query processing logic ### |
| 243 | |
| 244 | |
| 245 | async def process_query(context: ProcessingContext, query_text: str) -> float | None: |
| 246 | record = context.sr_index.get(query_text) |
| 247 | debug_context = searchlang.LanguageSearchDebugContext() |
| 248 | if context.debug1 == "skip" or context.debug2 == "skip": |
| 249 | if not record or ( |
| 250 | "searchQueryExpr" not in record or "compiledQueryExpr" not in record |
| 251 | ): |
| 252 | print("Can't skip stages 1 or 2, no precomputed outcomes found.") |
| 253 | else: |
| 254 | # Skipping stage 2 implies skipping stage 1, and we must supply the |
| 255 | # precomputed results for both stages. |
| 256 | debug_context.use_search_query = serialization.deserialize_object( |
| 257 | search_query_schema.SearchQuery, record["searchQueryExpr"] |
| 258 | ) |
| 259 | print("Skipping stage 1, substituting precomputed search query.") |
| 260 | if context.debug2 == "skip": |
| 261 | debug_context.use_compiled_search_query_exprs = ( |
| 262 | serialization.deserialize_object( |
| 263 | list[search.SearchQueryExpr], |
| 264 | record["compiledQueryExpr"], |
| 265 | ) |
| 266 | ) |
| 267 | print( |
| 268 | "Skipping stage 2, substituting precomputed compiled query expressions." |
| 269 | ) |
| 270 | prsep() |
| 271 | |
| 272 | result = await searchlang.search_conversation_with_language( |
| 273 | context.query_context.conversation, |
| 274 | context.query_translator, |
| 275 | query_text, |
| 276 | context.lang_search_options, |
| 277 | debug_context=debug_context, |
| 278 | ) |
| 279 | if isinstance(result, typechat.Failure): |
| 280 | print("Stages 1-3 failed:") |
| 281 | print(Fore.RED + str(result) + Fore.RESET) |
| 282 | return |
| 283 | search_results = result.value |
| 284 | |
| 285 | actual1 = debug_context.search_query |
| 286 | if actual1: |
| 287 | if context.debug1 == "full": |
| 288 | print("Stage 1 results:") |
| 289 | utils.pretty_print(actual1, Fore.GREEN, Fore.RESET) |
| 290 | prsep() |
| 291 | elif context.debug1 == "diff": |
| 292 | if record and "searchQueryExpr" in record: |
| 293 | print("Stage 1 diff:") |
| 294 | expected1 = serialization.deserialize_object( |
| 295 | search_query_schema.SearchQuery, record["searchQueryExpr"] |
| 296 | ) |
| 297 | compare_and_print_diff(expected1, actual1) |
| 298 | else: |
| 299 | print("Stage 1 diff unavailable") |
| 300 | prsep() |
| 301 | |
| 302 | actual2 = debug_context.search_query_expr |
| 303 | if actual2: |
| 304 | if context.debug2 == "full": |
| 305 | print("Stage 2 results:") |
| 306 | utils.pretty_print(actual2, Fore.GREEN, Fore.RESET) |
| 307 | prsep() |
| 308 | elif context.debug2 == "diff": |
| 309 | if record and "compiledQueryExpr" in record: |
| 310 | print("Stage 2 diff:") |
| 311 | expected2 = serialization.deserialize_object( |
| 312 | list[search.SearchQueryExpr], record["compiledQueryExpr"] |
| 313 | ) |
| 314 | compare_and_print_diff(expected2, actual2) |
| 315 | else: |
| 316 | print("Stage 2 diff unavailable") |
| 317 | prsep() |
| 318 | |
| 319 | actual3 = search_results |
| 320 | if context.debug3 == "full": |
| 321 | print("Stage 3 full results:") |
| 322 | utils.pretty_print(actual3, Fore.GREEN, Fore.RESET) |
| 323 | prsep() |
| 324 | elif context.debug3 == "nice": |
| 325 | print("Stage 3 nice results:") |
| 326 | for sr in search_results: |
| 327 | print_result(sr, context.query_context.conversation) |
| 328 | prsep() |
| 329 | elif context.debug3 == "diff": |
| 330 | if record and "results" in record: |
| 331 | print("Stage 3 diff:") |
| 332 | expected3: list[RawSearchResultData] = record["results"] |
| 333 | compare_results(expected3, actual3) |
| 334 | else: |
| 335 | print("Stage 3 diff unavailable") |
| 336 | prsep() |
| 337 | |
| 338 | all_answers, combined_answer = await answers.generate_answers( |
| 339 | context.answer_translator, |
| 340 | search_results, |
| 341 | context.query_context.conversation, |
| 342 | query_text, |
| 343 | options=context.answer_context_options, |
| 344 | ) |
| 345 | |
| 346 | if context.debug4 == "full": |
| 347 | utils.pretty_print(all_answers) |
| 348 | prsep() |
| 349 | if context.debug4 in ("full", "nice"): |
| 350 | if combined_answer.type == "NoAnswer": |
| 351 | print(Fore.RED + f"Failure: {combined_answer.whyNoAnswer}" + Fore.RESET) |
| 352 | else: |
| 353 | print(Fore.GREEN + f"{combined_answer.answer}" + Fore.RESET) |
| 354 | prsep() |
| 355 | elif context.debug4 == "diff": |
| 356 | if query_text in context.ar_index: |
| 357 | record = context.ar_index[query_text] |
| 358 | expected4: tuple[str, bool] = (record["answer"], not record["hasNoAnswer"]) |
| 359 | print("Stage 4 diff:") |
| 360 | match combined_answer.type: |
| 361 | case "NoAnswer": |
| 362 | actual4 = (combined_answer.whyNoAnswer or "", False) |
| 363 | case "Answered": |
| 364 | actual4 = (combined_answer.answer or "", True) |
| 365 | score = await compare_answers(context, expected4, actual4) |
| 366 | print(f"Score: {score:.3f}; Question: {query_text}") |
| 367 | return score |
| 368 | else: |
| 369 | print("Stage 4 diff unavailable; nice answer:") |
| 370 | if combined_answer.type == "NoAnswer": |
| 371 | print(Fore.RED + f"Failure: {combined_answer.whyNoAnswer}" + Fore.RESET) |
| 372 | else: |
| 373 | print(Fore.GREEN + f"{combined_answer.answer}" + Fore.RESET) |
| 374 | prsep() |
| 375 | |
| 376 | |
| 377 | def prsep(): |
| 378 | print("-" * 50) |
| 379 | |
| 380 | |
| 381 | ### CLI processing ### |
| 382 | |
| 383 | |
| 384 | def make_arg_parser(description: str) -> argparse.ArgumentParser: |
| 385 | line_width = utils.cap(144, shutil.get_terminal_size().columns) |
| 386 | parser = argparse.ArgumentParser( |
| 387 | description=description, |
| 388 | formatter_class=lambda *a, **b: argparse.HelpFormatter( |
| 389 | *a, **b, max_help_position=35 if line_width >= 100 else 28, width=line_width |
| 390 | ), |
| 391 | ) |
| 392 | |
| 393 | default_podcast_file = "testdata/Episode_53_AdrianTchaikovsky_index" |
| 394 | parser.add_argument( |
| 395 | "--podcast", |
| 396 | type=str, |
| 397 | default=default_podcast_file, |
| 398 | help="Path to the podcast index files (excluding the '_index.json' suffix)", |
| 399 | ) |
| 400 | default_qafile = "testdata/Episode_53_Answer_results.json" |
| 401 | explain_qa = "a list of questions and answers to test the full pipeline" |
| 402 | parser.add_argument( |
| 403 | "--qafile", |
| 404 | type=str, |
| 405 | default=default_qafile, |
| 406 | help=f"Path to the Answer_results.json file ({explain_qa})", |
| 407 | ) |
| 408 | default_srfile = "testdata/Episode_53_Search_results.json" |
| 409 | explain_sr = "a list of intermediate results from stages 1, 2 and 3" |
| 410 | parser.add_argument( |
| 411 | "--srfile", |
| 412 | type=str, |
| 413 | default=default_srfile, |
| 414 | help=f"Path to the Search_results.json file ({explain_sr})", |
| 415 | ) |
| 416 | |
| 417 | batch = parser.add_argument_group("Batch mode options") |
| 418 | batch.add_argument( |
| 419 | "--batch", |
| 420 | action="store_true", |
| 421 | help="Run in batch mode, suppressing interactive prompts.", |
| 422 | ) |
| 423 | batch.add_argument( |
| 424 | "--offset", |
| 425 | type=int, |
| 426 | default=0, |
| 427 | help="Number of initial Q/A pairs to skip (default none)", |
| 428 | ) |
| 429 | batch.add_argument( |
| 430 | "--limit", |
| 431 | type=int, |
| 432 | default=0, |
| 433 | help="Number of Q/A pairs to process (default all)", |
| 434 | ) |
| 435 | batch.add_argument( |
| 436 | "--start", |
| 437 | type=int, |
| 438 | default=0, |
| 439 | help="Do just this question (similar to --offset START-1 --limit 1)", |
| 440 | ) |
| 441 | |
| 442 | debug = parser.add_argument_group("Debug options") |
| 443 | debug.add_argument( |
| 444 | "--debug", |
| 445 | type=str, |
| 446 | default=None, |
| 447 | choices=["none", "diff", "full"], |
| 448 | help="Default debug level: 'none' for no debug output, 'diff' for diff output, " |
| 449 | "'full' for full debug output.", |
| 450 | ) |
| 451 | arg_helper = lambda key: typing.get_args(ProcessingContext.__annotations__[key]) |
| 452 | debug.add_argument( |
| 453 | "--debug1", |
| 454 | type=str, |
| 455 | default=None, |
| 456 | choices=arg_helper("debug1"), |
| 457 | help="Debug level override for stage 1: like --debug; or 'skip' to skip stage 1.", |
| 458 | ) |
| 459 | debug.add_argument( |
| 460 | "--debug2", |
| 461 | type=str, |
| 462 | default=None, |
| 463 | choices=arg_helper("debug2"), |
| 464 | help="Debug level override for stage 2: like --debug; or 'skip' to skip stages 1-2.", |
| 465 | ) |
| 466 | debug.add_argument( |
| 467 | "--debug3", |
| 468 | type=str, |
| 469 | default=None, |
| 470 | choices=arg_helper("debug3"), |
| 471 | help="Debug level override for stage 3: like --debug; or 'nice' to print answer only.", |
| 472 | ) |
| 473 | debug.add_argument( |
| 474 | "--debug4", |
| 475 | type=str, |
| 476 | default=None, |
| 477 | choices=arg_helper("debug4"), |
| 478 | help="Debug level override for stage 4: like --debug; or 'nice' to print answer only.", |
| 479 | ) |
| 480 | debug.add_argument( |
| 481 | "--alt-schema", |
| 482 | type=str, |
| 483 | default=None, |
| 484 | help="Path to alternate schema file for query translator (modifies stage 1).", |
| 485 | ) |
| 486 | debug.add_argument( |
| 487 | "--show-schema", |
| 488 | action="store_true", |
| 489 | help="Show the TypeScript schema computed by typechat.", |
| 490 | ) |
| 491 | |
| 492 | return parser |
| 493 | |
| 494 | |
| 495 | def fill_in_debug_defaults( |
| 496 | parser: argparse.ArgumentParser, args: argparse.Namespace |
| 497 | ) -> None: |
| 498 | # In batch mode, defaults are diff, diff, diff, diff. |
| 499 | # In interactive mode they are none, none, none, nice. |
| 500 | if not args.batch: |
| 501 | if args.start or args.offset or args.limit: |
| 502 | parser.exit(2, "Error: --start, --offset and --limit require --batch\n") |
| 503 | else: |
| 504 | if args.start: |
| 505 | if args.offset != 0: |
| 506 | parser.exit(2, "Error: --start and --offset can't be both set\n") |
| 507 | args.offset = args.start - 1 |
| 508 | if args.limit == 0: |
| 509 | args.limit = 1 |
| 510 | args.debug = args.debug or "diff" |
| 511 | |
| 512 | args.debug1 = args.debug1 or args.debug or "none" |
| 513 | args.debug2 = args.debug2 or args.debug or "none" |
| 514 | args.debug3 = args.debug3 or args.debug or "none" |
| 515 | args.debug4 = args.debug4 or args.debug or "nice" |
| 516 | if args.debug2 == "skip": |
| 517 | args.debug1 = "skip" # Skipping stage 2 implies skipping stage 1. |
| 518 | |
| 519 | |
| 520 | ### Data loading ### |
| 521 | |
| 522 | |
| 523 | def load_podcast_index( |
| 524 | podcast_file: str, settings: importing.ConversationSettings |
| 525 | ) -> query.QueryEvalContext: |
| 526 | with utils.timelog(f"load podcast from {podcast_file!r}"): |
| 527 | conversation = podcast.Podcast.read_from_file(podcast_file, settings) |
| 528 | assert conversation is not None, f"Failed to load podcast from {podcast_file!r}" |
| 529 | return query.QueryEvalContext(conversation) |
| 530 | |
| 531 | |
| 532 | def load_index_file[T: Mapping[str, typing.Any]]( |
| 533 | file: str, selector: str, cls: type[T] |
| 534 | ) -> tuple[list[T], dict[str, T]]: |
| 535 | # If this crashes, the file is malformed -- go figure it out. |
| 536 | try: |
| 537 | with open(file) as f: |
| 538 | lst: list[T] = json.load(f) |
| 539 | except FileNotFoundError as err: |
| 540 | print(Fore.RED + str(err) + Fore.RESET) |
| 541 | lst = [] |
| 542 | index = {item[selector]: item for item in lst} |
| 543 | if len(index) != len(lst): |
| 544 | print(f"{len(lst) - len(index)} duplicate items found in {file!r}. ") |
| 545 | return lst, index |
| 546 | |
| 547 | |
| 548 | ### Debug output ### |
| 549 | |
| 550 | |
| 551 | def print_result[TMessage: IMessage, TIndex: ITermToSemanticRefIndex]( |
| 552 | result: search.ConversationSearchResult, |
| 553 | conversation: IConversation[TMessage, TIndex], |
| 554 | ) -> None: |
| 555 | print( |
| 556 | f"Raw query: {result.raw_query_text};", |
| 557 | f"{len(result.message_matches)} message matches,", |
| 558 | f"{len(result.knowledge_matches)} knowledge matches", |
| 559 | ) |
| 560 | if result.message_matches: |
| 561 | print("Message matches:") |
| 562 | for scored_ord in sorted( |
| 563 | result.message_matches, key=lambda x: x.score, reverse=True |
| 564 | ): |
| 565 | score = scored_ord.score |
| 566 | msg_ord = scored_ord.message_ordinal |
| 567 | msg = conversation.messages[msg_ord] |
| 568 | assert msg.metadata is not None # For type checkers |
| 569 | text = " ".join(msg.text_chunks).strip() |
| 570 | print( |
| 571 | f"({score:5.1f}) M={msg_ord:d}: " |
| 572 | f"{msg.metadata.source!s:>15.15s}: " |
| 573 | f"{repr(text)[1:-1]:<150.150s} " |
| 574 | ) |
| 575 | if result.knowledge_matches: |
| 576 | print(f"Knowledge matches ({', '.join(sorted(result.knowledge_matches))}):") |
| 577 | for key, value in sorted(result.knowledge_matches.items()): |
| 578 | print(f"Type {key} -- {value.term_matches}:") |
| 579 | for scored_sem_ref_ord in value.semantic_ref_matches: |
| 580 | score = scored_sem_ref_ord.score |
| 581 | sem_ref_ord = scored_sem_ref_ord.semantic_ref_ordinal |
| 582 | if conversation.semantic_refs is None: |
| 583 | print(f" Ord: {sem_ref_ord} (score {score})") |
| 584 | else: |
| 585 | sem_ref = conversation.semantic_refs[sem_ref_ord] |
| 586 | msg_ord = sem_ref.range.start.message_ordinal |
| 587 | chunk_ord = sem_ref.range.start.chunk_ordinal |
| 588 | msg = conversation.messages[msg_ord] |
| 589 | print( |
| 590 | f"({score:5.1f}) M={msg_ord}: " |
| 591 | f"S={summarize_knowledge(sem_ref)}" |
| 592 | ) |
| 593 | |
| 594 | |
| 595 | def summarize_knowledge(sem_ref: SemanticRef) -> str: |
| 596 | """Summarize the knowledge in a SemanticRef.""" |
| 597 | knowledge = sem_ref.knowledge |
| 598 | if knowledge is None: |
| 599 | return f"{sem_ref.semantic_ref_ordinal}: <No knowledge>" |
| 600 | match sem_ref.knowledge_type: |
| 601 | case "entity": |
| 602 | entity = knowledge |
| 603 | assert isinstance(entity, kplib.ConcreteEntity) |
| 604 | res = [f"{entity.name} [{', '.join(entity.type)}]"] |
| 605 | if entity.facets: |
| 606 | for facet in entity.facets: |
| 607 | value = facet.value |
| 608 | if isinstance(value, kplib.Quantity): |
| 609 | value = f"{value.amount} {value.units}" |
| 610 | elif isinstance(value, float) and value.is_integer(): |
| 611 | value = int(value) |
| 612 | res.append(f"<{facet.name}:{value}>") |
| 613 | return f"{sem_ref.semantic_ref_ordinal}: {' '.join(res)}" |
| 614 | case "action": |
| 615 | action = knowledge |
| 616 | assert isinstance(action, kplib.Action) |
| 617 | res = [] |
| 618 | res.append("/".join(repr(verb) for verb in action.verbs)) |
| 619 | if action.verb_tense: |
| 620 | res.append(f"[{action.verb_tense}]") |
| 621 | if action.subject_entity_name != "none": |
| 622 | res.append(f"subj={action.subject_entity_name!r}") |
| 623 | if action.object_entity_name != "none": |
| 624 | res.append(f"obj={action.object_entity_name!r}") |
| 625 | if action.indirect_object_entity_name != "none": |
| 626 | res.append(f"ind_obj={action.indirect_object_entity_name}") |
| 627 | if action.params: |
| 628 | for param in action.params: |
| 629 | if isinstance(param, kplib.ActionParam): |
| 630 | res.append(f"<{param.name}:{param.value}>") |
| 631 | else: |
| 632 | res.append(f"<{param}>") |
| 633 | if action.subject_entity_facet is not None: |
| 634 | res.append(f"subj_facet={action.subject_entity_facet}") |
| 635 | return f"{sem_ref.semantic_ref_ordinal}: {' '.join(res)}" |
| 636 | case "topic": |
| 637 | topic = knowledge |
| 638 | assert isinstance(topic, Topic) |
| 639 | return f"{sem_ref.semantic_ref_ordinal}: {topic.text!r}]" |
| 640 | case "tag": |
| 641 | tag = knowledge |
| 642 | assert isinstance(tag, str) |
| 643 | return f"{sem_ref.semantic_ref_ordinal}: #{tag!r}" |
| 644 | case _: |
| 645 | return f"{sem_ref.semantic_ref_ordinal}: {sem_ref.knowledge!r}" |
| 646 | |
| 647 | |
| 648 | def compare_results( |
| 649 | matches_records: list[RawSearchResultData], |
| 650 | results: list[search.ConversationSearchResult], |
| 651 | ) -> bool: |
| 652 | if len(results) != len(matches_records): |
| 653 | print(f"(Result sizes mismatch, {len(results)} != {len(matches_records)})") |
| 654 | return False |
| 655 | res = True |
| 656 | for result, record in zip(results, matches_records): |
| 657 | if not compare_message_ordinals( |
| 658 | result.message_matches, record["messageMatches"] |
| 659 | ): |
| 660 | res = False |
| 661 | if not compare_semantic_ref_ordinals( |
| 662 | ( |
| 663 | [] |
| 664 | if "entity" not in result.knowledge_matches |
| 665 | else result.knowledge_matches["entity"].semantic_ref_matches |
| 666 | ), |
| 667 | record.get("entityMatches", []), |
| 668 | "entity", |
| 669 | ): |
| 670 | res = False |
| 671 | if not compare_semantic_ref_ordinals( |
| 672 | ( |
| 673 | [] |
| 674 | if "action" not in result.knowledge_matches |
| 675 | else result.knowledge_matches["action"].semantic_ref_matches |
| 676 | ), |
| 677 | record.get("actionMatches", []), |
| 678 | "action", |
| 679 | ): |
| 680 | res = False |
| 681 | if not compare_semantic_ref_ordinals( |
| 682 | ( |
| 683 | [] |
| 684 | if "topic" not in result.knowledge_matches |
| 685 | else result.knowledge_matches["topic"].semantic_ref_matches |
| 686 | ), |
| 687 | record.get("topicMatches", []), |
| 688 | "topic", |
| 689 | ): |
| 690 | res = False |
| 691 | return res |
| 692 | |
| 693 | |
| 694 | # Special case: In the Podcast, these messages are all Kevin saying "Yeah", |
| 695 | # so if the difference is limited to these, we consider it a match. |
| 696 | NOISE_MESSAGES = frozenset({42, 46, 52, 68, 70}) |
| 697 | |
| 698 | |
| 699 | def compare_message_ordinals(aa: list[ScoredMessageOrdinal], b: list[int]) -> bool: |
| 700 | a = [aai.message_ordinal for aai in aa] |
| 701 | if set(a) ^ set(b) <= NOISE_MESSAGES: |
| 702 | return True |
| 703 | print("Message ordinals do not match:") |
| 704 | utils.list_diff(" Expected:", b, " Actual:", a, max_items=20) |
| 705 | return False |
| 706 | |
| 707 | |
| 708 | def compare_semantic_ref_ordinals( |
| 709 | aa: list[ScoredSemanticRefOrdinal], b: list[int], label: str |
| 710 | ) -> bool: |
| 711 | a = [aai.semantic_ref_ordinal for aai in aa] |
| 712 | if sorted(a) == sorted(b): |
| 713 | return True |
| 714 | print(f"{label.capitalize()} SemanticRef ordinals do not match:") |
| 715 | utils.list_diff(" Expected:", b, " Actual:", a, max_items=20) |
| 716 | return False |
| 717 | |
| 718 | |
| 719 | def compare_and_print_diff(a: object, b: object) -> bool: # True if equal |
| 720 | """Diff two objects whose repr() is a valid Python expression.""" |
| 721 | if a == b: |
| 722 | return True |
| 723 | a_repr = repr(a) |
| 724 | b_repr = repr(b) |
| 725 | if a_repr == b_repr: |
| 726 | return True |
| 727 | # Shorten floats so slight differences in score etc. don't cause false positives. |
| 728 | a_repr = re.sub(r"\b\d\.\d\d+", lambda m: f"{float(m.group()):.3f}", a_repr) |
| 729 | b_repr = re.sub(r"\b\d\.\d\d+", lambda m: f"{float(m.group()):.3f}", b_repr) |
| 730 | if a_repr == b_repr: |
| 731 | return True |
| 732 | a_formatted = utils.format_code(a_repr) |
| 733 | b_formatted = utils.format_code(b_repr) |
| 734 | print_diff(a_formatted, b_formatted, n=2) |
| 735 | return False |
| 736 | |
| 737 | |
| 738 | async def compare_answers( |
| 739 | context: ProcessingContext, expected: tuple[str, bool], actual: tuple[str, bool] |
| 740 | ) -> float: |
| 741 | expected_text, expected_success = expected |
| 742 | actual_text, actual_success = actual |
| 743 | |
| 744 | if expected_success != actual_success: |
| 745 | print(f"Expected success: {expected_success}; actual: {actual_success}") |
| 746 | return 0.000 |
| 747 | |
| 748 | if not actual_success: |
| 749 | print(Fore.GREEN + f"Both failed" + Fore.RESET) |
| 750 | return 1.001 |
| 751 | |
| 752 | if expected_text == actual_text: |
| 753 | print(Fore.GREEN + f"Both equal" + Fore.RESET) |
| 754 | return 1.000 |
| 755 | |
| 756 | if len(expected_text.splitlines()) <= 100 and len(actual_text.splitlines()) <= 100: |
| 757 | n = 100 |
| 758 | else: |
| 759 | n = 2 |
| 760 | print_diff(expected_text, actual_text, n=n) |
| 761 | return await equality_score(context, expected_text, actual_text) |
| 762 | |
| 763 | |
| 764 | def print_diff(a: str, b: str, n: int) -> None: |
| 765 | diff = difflib.unified_diff( |
| 766 | a.splitlines(), |
| 767 | b.splitlines(), |
| 768 | fromfile="expected", |
| 769 | tofile="actual", |
| 770 | n=n, |
| 771 | ) |
| 772 | for x in diff: |
| 773 | if x.startswith("-"): |
| 774 | print(Fore.RED + x.rstrip("\n") + Fore.RESET) |
| 775 | elif x.startswith("+"): |
| 776 | print(Fore.GREEN + x.rstrip("\n") + Fore.RESET) |
| 777 | else: |
| 778 | print(x.rstrip("\n")) |
| 779 | |
| 780 | |
| 781 | async def equality_score(context: ProcessingContext, a: str, b: str) -> float: |
| 782 | if a == b: |
| 783 | return 1.0 |
| 784 | if a.lower() == b.lower(): |
| 785 | return 0.999 |
| 786 | embeddings = await context.embedding_model.get_embeddings([a, b]) |
| 787 | assert embeddings.shape[0] == 2, "Expected two embeddings" |
| 788 | return np.dot(embeddings[0], embeddings[1]) |
| 789 | |
| 790 | |
| 791 | ### Run main ### |
| 792 | |
| 793 | if __name__ == "__main__": |
| 794 | try: |
| 795 | main() |
| 796 | except (KeyboardInterrupt, BrokenPipeError): |
| 797 | print() |
| 798 | sys.exit(1) |
| 799 | |