File size: 62,701 Bytes
da605e9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 | """
Goon analysis agent β all modules in one file.
Sections:
1. Response formatter
2. Data inspection & sampling
3. Question router
4. Image analysis
5. Text pattern extraction
6. Analysis execution (count, trend, stats, search, word freq, compare)
7. Core agent loop
"""
from __future__ import annotations
# ββ stdlib βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
import base64
import glob
import io
import json
import os
import re
import traceback as _traceback
import urllib.request
from dataclasses import dataclass
from datetime import datetime, timezone
from pathlib import Path
# ββ third-party ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
import anthropic
import openai # Together AI uses an OpenAI-compatible endpoint
import pandas as pd
import plotly.express as px
import plotly.io as pio
import pyarrow.dataset as ds
import pyarrow.parquet as pq
from pyarrow.compute import field
from sklearn.metrics import cohen_kappa_score
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 1. Response formatter
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def format_result(result: dict, answer_text: str = "") -> str:
"""Combine Claude's prose answer with the structured analysis result as markdown."""
lines = []
if answer_text:
lines.append("## Answer\n")
lines.append(answer_text.strip())
lines.append("")
dataset = result.get("dataset", "")
if dataset:
lines.append("## What was analysed\n")
lines.append(f"- Dataset: `{dataset}`")
if result.get("subreddit_filter"):
lines.append(f"- Subreddit filter: `{result['subreddit_filter']}`")
if result.get("group_col"):
lines.append(f"- Grouped by: `{result['group_col']}`")
if result.get("value_col"):
lines.append(f"- Value column: `{result['value_col']}`")
lines.append("")
table = result.get("table")
if table:
lines.append("## Results\n")
lines.append(_dict_list_to_md_table(table))
lines.append("")
saved = [result[k] for k in ("saved_csv", "saved_png") if result.get(k)]
if saved:
lines.append("## Saved outputs\n")
for s in saved:
lines.append(f"- `{s}`")
lines.append("")
return "\n".join(lines)
def _dict_list_to_md_table(records: list[dict]) -> str:
if not records:
return "_No results._"
headers = list(records[0].keys())
rows = [[str(r.get(h, "")) for h in headers] for r in records]
widths = [max(len(h), max((len(r[i]) for r in rows), default=0)) for i, h in enumerate(headers)]
sep = "| " + " | ".join("-" * w for w in widths) + " |"
header_row = "| " + " | ".join(h.ljust(widths[i]) for i, h in enumerate(headers)) + " |"
data_rows = [
"| " + " | ".join(cell.ljust(widths[i]) for i, cell in enumerate(row)) + " |"
for row in rows[:50]
]
return "\n".join([header_row, sep] + data_rows)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 2. Data inspection & sampling
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
DATA_DIR = Path(os.environ.get("DATA_DIR", Path(__file__).parent.parent / "data"))
OUTPUTS_DIR = Path(__file__).parent / "outputs"
OUTPUTS_DIR.mkdir(exist_ok=True)
METADATA_CACHE = OUTPUTS_DIR / "dataset_metadata.json"
def _best(name: str) -> Path:
"""Prefer the full rebuilt parquet over the original, but validate it first."""
full = DATA_DIR / f"{name}_full.parquet"
orig = DATA_DIR / f"{name}.parquet"
if full.exists():
try:
pq.read_schema(full)
return full
except Exception:
pass
return orig
DATASETS = {
"posts": _best("posts"),
"comments": _best("comments"),
"corpus_clean": DATA_DIR / "corpus_clean.parquet",
"titles": _best("titles"),
}
def _dataset_path(name: str) -> Path:
path = DATASETS.get(name)
if path is None or not path.exists():
raise FileNotFoundError(f"Dataset '{name}' not found at {path}")
return path
def _load(name: str, columns: list[str] | None = None) -> pd.DataFrame:
return pd.read_parquet(_dataset_path(name), columns=columns)
def _scanner(name: str, columns: list[str] | None = None, filters: dict | None = None) -> ds.Scanner:
path = _dataset_path(name)
dataset = ds.dataset(path, format="parquet")
expression = None
for col, value in (filters or {}).items():
if col not in dataset.schema.names or value in (None, ""):
continue
clause = field(col) == value
expression = clause if expression is None else expression & clause
return dataset.scanner(columns=columns, filter=expression)
def _read_distinct_values(name: str, column: str, limit: int = 200) -> list[str] | None:
if column not in _schema_names(name):
return None
table = _scanner(name, columns=[column]).to_table()
values = table.column(column).drop_null().unique().to_pylist()
return sorted(str(v) for v in values)[:limit]
def _read_date_range(name: str) -> dict | None:
if "created_utc" not in _schema_names(name):
return None
table = _scanner(name, columns=["created_utc"]).to_table()
if table.num_rows == 0:
return None
series = table.column("created_utc").to_pandas().dropna()
if series.empty:
return None
return {
"earliest": datetime.fromtimestamp(series.min(), tz=timezone.utc).strftime("%Y-%m-%d"),
"latest": datetime.fromtimestamp(series.max(), tz=timezone.utc).strftime("%Y-%m-%d"),
}
def _schema_names(name: str) -> list[str]:
return pq.read_schema(_dataset_path(name)).names
def compute_dataset_metadata() -> dict:
result = {}
for name, path in DATASETS.items():
if not path.exists():
result[name] = {"available": False}
continue
parquet = pq.ParquetFile(path)
schema = parquet.schema_arrow
result[name] = {
"available": True,
"path": str(path),
"rows": parquet.metadata.num_rows,
"columns": {f.name: str(f.type) for f in schema},
"subreddits": _read_distinct_values(name, "subreddit"),
"date_range": _read_date_range(name),
"metadata_cached_at": datetime.now(timezone.utc).isoformat(),
}
METADATA_CACHE.write_text(json.dumps(result, indent=2))
return result
def get_dataset_metadata(refresh: bool = False) -> dict:
if METADATA_CACHE.exists() and not refresh:
return json.loads(METADATA_CACHE.read_text())
return compute_dataset_metadata()
def list_datasets(refresh: bool = False) -> dict:
"""Return cached dataset metadata instead of loading full tables."""
return get_dataset_metadata(refresh=refresh)
def sample_rows(
dataset: str,
n: int = 5,
filters: dict | None = None,
columns: list[str] | None = None,
) -> dict:
"""Return a small deterministic preview of rows from a dataset, optionally filtered."""
selected_columns = columns or _schema_names(dataset)
table = _scanner(dataset, columns=selected_columns, filters=filters).head(n)
df = table.to_pandas() if table.num_rows else pd.DataFrame(columns=selected_columns)
return {
"dataset": dataset,
"filters": filters or {},
"n_returned": len(df),
"rows": df.fillna("").to_dict(orient="records"),
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 3. Question router
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@dataclass(frozen=True)
class RoutePlan:
mode: str
allowed_tools: list[str]
guidance: str
ALL_TOOL_NAMES = [
"list_datasets", "sample_rows", "count_by_group", "trend_over_time",
"summary_stats", "top_posts", "text_search", "word_freq", "compare_groups",
"extract_frequency_patterns", "extract_dominance_patterns", "analyze_image_sample",
"export_reliability_sample", "compute_reliability",
]
def route_question(question: str) -> RoutePlan:
q = question.lower()
if any(t in q for t in ["image", "images", "photo", "photos", "visual", "depicted"]):
return RoutePlan(
mode="image",
allowed_tools=["list_datasets", "sample_rows", "analyze_image_sample", "export_reliability_sample", "compute_reliability"],
guidance="This is a visual-content question. Prefer image analysis tools and avoid text-only proxies. Always provide a coding_scheme.",
)
if any(t in q for t in ["reliability", "kappa", "human coding", "inter-rater", "validate"]):
return RoutePlan(
mode="reliability",
allowed_tools=["export_reliability_sample", "compute_reliability"],
guidance="This is a reliability/validation question. Use export_reliability_sample then compute_reliability.",
)
if any(t in q for t in ["how often", "how long", "times per", "every day", "session length", "streak"]):
return RoutePlan(
mode="pattern_frequency",
allowed_tools=["list_datasets", "sample_rows", "extract_frequency_patterns", "text_search"],
guidance="This is a behavioral frequency/duration question. Prefer regex pattern extraction over generic word counts.",
)
if any(t in q for t in ["dominant", "subordinate", "mistress", "goddess", "femdom", "submissive"]):
return RoutePlan(
mode="pattern_dominance",
allowed_tools=["list_datasets", "sample_rows", "extract_dominance_patterns", "text_search", "analyze_image_sample"],
guidance="This is a dominance/subordination framing question. Use the text pattern tool unless the user explicitly asks about images.",
)
if any(t in q for t in ["over time", "trend", "changed", "change over time", "monthly", "yearly"]):
return RoutePlan(
mode="trend",
allowed_tools=["list_datasets", "sample_rows", "trend_over_time", "count_by_group"],
guidance="This is a time-series question. Prefer trend_over_time and only use grouping/count tools to contextualize it.",
)
if any(t in q for t in ["compare", "difference", "versus", "vs", "higher", "lower"]):
return RoutePlan(
mode="compare",
allowed_tools=["list_datasets", "sample_rows", "compare_groups", "summary_stats", "count_by_group"],
guidance="This is a comparison question. Prefer compare_groups or summary_stats with explicit filters.",
)
if any(t in q for t in ["top", "highest", "best scoring", "most upvoted"]):
return RoutePlan(
mode="ranking",
allowed_tools=["list_datasets", "sample_rows", "top_posts", "summary_stats"],
guidance="This is a ranking question. Prefer top_posts and use summary_stats only if it supports the answer.",
)
if any(t in q for t in ["search", "find", "mention", "contains", "where people say"]):
return RoutePlan(
mode="search",
allowed_tools=["list_datasets", "sample_rows", "text_search", "top_posts"],
guidance="This is a retrieval question. Prefer text_search with the right dataset and text column.",
)
if any(t in q for t in ["common words", "most common words", "word frequency", "tokens"]):
return RoutePlan(
mode="lexical",
allowed_tools=["list_datasets", "sample_rows", "word_freq", "text_search"],
guidance="This is a lexical summary question. Prefer word_freq and inspect text samples only if needed.",
)
if any(t in q for t in ["how many", "count", "number of", "what proportion"]):
return RoutePlan(
mode="describe",
allowed_tools=["list_datasets", "sample_rows", "count_by_group", "summary_stats", "trend_over_time"],
guidance="This is a descriptive count question. Prefer count_by_group or summary_stats and keep the plan minimal.",
)
return RoutePlan(
mode="unknown",
allowed_tools=ALL_TOOL_NAMES,
guidance="Question type is ambiguous. Inspect metadata first, then choose the minimum reliable tool path.",
)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 4. Image analysis
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
VISION_MODEL = "Qwen/Qwen2-VL-72B-Instruct"
TOGETHER_BASE_URL = "https://api.together.xyz/v1"
DIRECT_IMAGE_DOMAINS = {"i.redd.it", "i.imgur.com", "i.redgifs.com"}
def _load_image_urls(subreddit: str | None = None, n: int = 50) -> pd.DataFrame:
pattern = str(DATA_DIR / "*_submissions_*.csv")
files = sorted(glob.glob(pattern))
if subreddit:
files = [f for f in files if Path(f).name.lower().startswith(subreddit.lower())]
needed_cols = ["subreddit", "title", "url", "domain", "score", "is_self"]
frames = []
for f in files:
try:
df = pd.read_csv(f, usecols=lambda c: c in needed_cols, low_memory=False)
if "is_self" in df.columns:
df = df[df["is_self"] == False]
if "url" in df.columns and "domain" in df.columns:
df = df[df["domain"].isin(DIRECT_IMAGE_DOMAINS)].dropna(subset=["url"])
frames.append(df[["subreddit", "title", "url", "domain", "score"]])
except Exception:
continue
if not frames:
return pd.DataFrame()
combined = pd.concat(frames, ignore_index=True)
if len(combined) > n * 10:
combined = combined.sample(min(n * 10, len(combined)), random_state=42)
return combined.head(n * 10)
def _fetch_image_b64(url: str, timeout: int = 8) -> tuple[str, str] | None:
try:
url.encode("ascii")
except UnicodeEncodeError:
return None
try:
req = urllib.request.Request(url, headers={"User-Agent": "Mozilla/5.0 (research bot)"})
with urllib.request.urlopen(req, timeout=timeout) as resp:
content_type = resp.headers.get("Content-Type", "image/jpeg").split(";")[0].strip()
if not content_type.startswith("image/"):
return None
data = resp.read()
if len(data) < 1000:
return None
return base64.standard_b64encode(data).decode("utf-8"), content_type
except Exception:
return None
def analyze_image_sample(
question: str,
subreddit: str | None = None,
n_sample: int = 100,
coding_scheme: dict | None = None,
) -> dict:
"""
Sample image posts, fetch them, and ask Qwen2-VL a structured content-analysis question.
Uses Together AI (no content filters). n_sample is uncapped β set as needed.
"""
client = openai.OpenAI(
api_key=os.environ["TOGETHER_API_KEY"],
base_url=TOGETHER_BASE_URL,
)
candidates = _load_image_urls(subreddit=subreddit, n=n_sample * 5)
if candidates.empty:
return {
"analysis": "analyze_image_sample",
"error": "No direct image URLs found in raw CSVs for the given filters.",
"subreddit_filter": subreddit,
}
if coding_scheme:
scheme_text = "\n".join(f"- {k}: {v}" for k, v in coding_scheme.items())
prompt = (
f"{question}\n\nCoding scheme:\n{scheme_text}\n\n"
"Reply with ONLY the label and a one-sentence justification, "
"formatted as: LABEL | justification"
)
else:
prompt = (
f"{question}\n\n"
"Reply with a short structured answer. "
"If you cannot determine this from the image, reply: UNCLEAR | reason"
)
results = []
attempted = 0
for _, row in candidates.iterrows():
if len(results) >= n_sample:
break
attempted += 1
img = _fetch_image_b64(row["url"])
if img is None:
continue
b64data, media_type = img
try:
response = client.chat.completions.create(
model=VISION_MODEL,
max_tokens=200,
messages=[{
"role": "user",
"content": [
{"type": "image_url", "image_url": {"url": f"data:{media_type};base64,{b64data}"}},
{"type": "text", "text": prompt},
],
}],
)
answer = response.choices[0].message.content.strip()
except Exception as e:
answer = f"ERROR | {e}"
parts = answer.split("|", 1)
label = parts[0].strip().upper() if parts else "UNCLEAR"
justification = parts[1].strip() if len(parts) > 1 else ""
def _ascii_safe(s: str) -> str:
return s.encode("ascii", errors="replace").decode("ascii")
results.append({
"subreddit": _ascii_safe(str(row.get("subreddit", ""))),
"title": _ascii_safe(str(row.get("title", ""))),
"url": row["url"],
"label": label,
"justification": _ascii_safe(justification),
"score": row.get("score", None),
})
label_counts: dict[str, int] = {}
for r in results:
label_counts[r["label"]] = label_counts.get(r["label"], 0) + 1
total_coded = len(results)
saved_csv = None
if results:
out_df = pd.DataFrame(results)
stem = f"image_analysis_{subreddit or 'all'}"
saved_csv = str(OUTPUTS_DIR / f"{stem}.csv")
out_df.to_csv(saved_csv, index=False)
return {
"analysis": "analyze_image_sample",
"question": question,
"subreddit_filter": subreddit,
"n_attempted": attempted,
"n_successfully_coded": total_coded,
"label_counts": label_counts,
"label_pct": {k: round(v / total_coded * 100, 1) for k, v in label_counts.items()} if total_coded else {},
"per_image_results": results,
"saved_csv": saved_csv,
"caveats": [
"Sample limited to direct-image domains (i.redd.it, i.imgur.com, i.redgifs.com) β galleries and videos excluded.",
f"Vision model: {VISION_MODEL} via Together AI.",
"Coded by a single model β validate with human reliability sample before reporting.",
],
}
def export_reliability_sample(
source_csv: str | None = None,
n: int = 200,
random_state: int = 42,
) -> dict:
"""
Draw a stratified random sample of n images from a completed image_analysis CSV
for human coding. Saves a CSV with an empty human_label column.
"""
if source_csv is None:
# Default to most recently written image_analysis file
candidates = sorted(OUTPUTS_DIR.glob("image_analysis_*.csv"))
if not candidates:
return {"error": "No image_analysis CSV found in outputs/. Run analyze_image_sample first."}
source_csv = str(candidates[-1])
df = pd.read_csv(source_csv)
df = df[df["label"].notna() & ~df["label"].str.startswith("ERROR")]
# Stratified sample by label
sampled = (
df.groupby("label", group_keys=False)
.apply(lambda g: g.sample(min(len(g), max(1, int(n * len(g) / len(df)))), random_state=random_state))
)
# Top up to exactly n if rounding left us short
if len(sampled) < n and len(df) >= n:
remaining = df[~df.index.isin(sampled.index)]
top_up = remaining.sample(n - len(sampled), random_state=random_state)
sampled = pd.concat([sampled, top_up])
sampled = sampled.sample(frac=1, random_state=random_state).reset_index(drop=True) # shuffle
sampled.insert(0, "image_id", range(1, len(sampled) + 1))
sampled = sampled.rename(columns={"label": "model_label", "justification": "model_justification"})
sampled["human_label"] = ""
out_cols = ["image_id", "url", "title", "subreddit", "model_label", "model_justification", "human_label"]
out_cols = [c for c in out_cols if c in sampled.columns]
out_path = str(OUTPUTS_DIR / "reliability_sample.csv")
sampled[out_cols].to_csv(out_path, index=False)
return {
"analysis": "export_reliability_sample",
"source_csv": source_csv,
"n_exported": len(sampled),
"label_distribution": sampled["model_label"].value_counts().to_dict(),
"saved_csv": out_path,
"next_step": "Fill in the human_label column, then run compute_reliability.",
}
def compute_reliability(human_csv_path: str | None = None) -> dict:
"""
Compute Cohen's kappa between model_label and human_label columns
in a completed reliability sample CSV.
"""
if human_csv_path is None:
human_csv_path = str(OUTPUTS_DIR / "reliability_sample.csv")
df = pd.read_csv(human_csv_path)
df = df[df["human_label"].notna() & (df["human_label"].astype(str).str.strip() != "")]
if len(df) < 2:
return {"error": "Not enough human-coded rows. Fill in human_label column first."}
model = df["model_label"].astype(str).str.strip().str.upper()
human = df["human_label"].astype(str).str.strip().str.upper()
kappa = cohen_kappa_score(human, model)
pct_agreement = round((human == model).mean() * 100, 1)
per_label = {}
for label in sorted(human.unique()):
h = (human == label)
m = (model == label)
tp = int((h & m).sum())
fp = int((~h & m).sum())
fn = int((h & ~m).sum())
per_label[label] = {"human_n": int(h.sum()), "model_n": int(m.sum()),
"exact_matches": tp, "false_positives": fp, "false_negatives": fn}
report = {
"analysis": "compute_reliability",
"n_coded": len(df),
"cohens_kappa": round(kappa, 3),
"percent_agreement": pct_agreement,
"interpretation": (
"excellent (ΞΊ β₯ 0.80)" if kappa >= 0.80 else
"substantial (ΞΊ 0.60β0.79)" if kappa >= 0.60 else
"moderate (ΞΊ 0.40β0.59)" if kappa >= 0.40 else
"fair (ΞΊ 0.20β0.39)" if kappa >= 0.20 else
"poor (ΞΊ < 0.20)"
),
"per_label": per_label,
}
out_path = str(OUTPUTS_DIR / "reliability_report.json")
Path(out_path).write_text(json.dumps(report, indent=2))
report["saved_json"] = out_path
return report
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 5. Text pattern extraction
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
FREQUENCY_PATTERNS = {
"times_per_day": [
r"\b(\d+)\s*(?:times?|x)\s*(?:a|per)\s*day\b",
r"\b(\d+)\s*(?:times?|x)\s*daily\b",
],
"times_per_week": [
r"\b(\d+)\s*(?:times?|x)\s*(?:a|per)\s*week\b",
r"\b(\d+)\s*(?:times?|x)\s*weekly\b",
],
"hours_per_session": [
r"\b(\d+(?:\.\d+)?)\s*(?:hours?|hrs?)\b",
r"\b(\d+(?:\.\d+)?)\s*(?:hours?|hrs?)\s*(?:session|goon|long|straight|solid|non.?stop)\b",
],
"all_day": [
r"\ball\s*day\b", r"\ball\s*night\b", r"\ball\s*weekend\b", r"\bfor\s*hours\b",
],
"daily_habit": [
r"\bevery\s*day\b", r"\bevery\s*night\b", r"\bdaily\b", r"\bmost\s*days?\b",
],
"streak_days": [
r"\b(\d+)\s*(?:days?\s*(?:in\s*a\s*row|straight|streak|running))\b",
r"\b(\d+)\s*(?:-|β)?\s*day\s*(?:streak|binge)\b",
],
}
DOMINANCE_PATTERNS = {
"dominant_language": [
r"\bdominat(?:e|es|ed|ing|ion|rix|rix)\b", r"\bfem(?:dom|domme)\b",
r"\bmistress\b", r"\bgoddess\b", r"\bqueen\b", r"\bowner\b",
r"\balpha\b", r"\bin\s*control\b", r"\bboss\b",
],
"subordinate_language": [
r"\bsubmiss(?:ive|ion)\b", r"\bsub\b", r"\bobedient\b", r"\bslave\b",
r"\bpet\b", r"\bslut\b", r"\bwhore\b", r"\bused\b",
r"\bcontrolled\b", r"\bworshiped?\b", r"\bworship(?:ped|ing)\b",
],
"neutral_object": [
r"\bperfect\b", r"\bbeautiful\b", r"\bhot\b", r"\bsexy\b", r"\bstunning\b",
],
}
def _compile(patterns: list[str]) -> re.Pattern:
return re.compile("|".join(patterns), re.IGNORECASE)
def extract_frequency_patterns(
dataset: str = "comments",
text_col: str = "body",
subreddit: str | None = None,
n_examples: int = 5,
sample_size: int = 5_000_000,
) -> dict:
"""Mine text for frequency and duration language across the full dataset."""
cols = [text_col] + (["subreddit"] if subreddit else [])
df = _scanner(
dataset, columns=cols,
filters={"subreddit": subreddit} if subreddit else None,
).head(sample_size).to_pandas()
text = df[text_col].fillna("")
total_docs = len(text)
results = {}
for category, pats in FREQUENCY_PATTERNS.items():
regex = _compile(pats)
matches_mask = text.str.contains(regex.pattern, regex=True, na=False)
hit_texts = text[matches_mask]
values = []
for pat in pats:
r = re.compile(pat, re.IGNORECASE)
for t in hit_texts:
for m in r.finditer(t):
if m.groups():
try:
values.append(float(m.group(1)))
except (IndexError, ValueError):
pass
raw_examples = hit_texts.sample(min(n_examples, len(hit_texts)), random_state=42).tolist() if len(hit_texts) > 0 else []
results[category] = {
"count": int(matches_mask.sum()),
"pct_of_docs": round(matches_mask.mean() * 100, 3),
"numeric_values": sorted(values)[:50] if values else [],
"mean_value": round(sum(values) / len(values), 2) if values else None,
"examples": [t.encode("ascii", errors="replace").decode("ascii") for t in raw_examples],
}
return {
"analysis": "extract_frequency_patterns",
"dataset": dataset,
"text_col": text_col,
"subreddit_filter": subreddit,
"total_docs_sampled": total_docs,
"patterns": results,
}
def extract_dominance_patterns(
dataset: str = "comments",
text_col: str = "body",
subreddit: str | None = None,
sample_size: int = 5_000_000,
) -> dict:
"""Count dominant, subordinate, and neutral language in text."""
cols = [text_col] + (["subreddit"] if subreddit else [])
df = _scanner(
dataset, columns=cols,
filters={"subreddit": subreddit} if subreddit else None,
).head(sample_size).to_pandas()
text = df[text_col].fillna("")
total_docs = len(text)
results = {}
for category, pats in DOMINANCE_PATTERNS.items():
regex = _compile(pats)
mask = text.str.contains(regex, na=False)
hits = text[mask]
raw_examples = hits.sample(min(5, len(hits)), random_state=42).tolist() if len(hits) > 0 else []
results[category] = {
"count": int(mask.sum()),
"pct_of_docs": round(mask.mean() * 100, 3),
"examples": [t.encode("ascii", errors="replace").decode("ascii") for t in raw_examples],
}
dom = results.get("dominant_language", {}).get("count", 0)
sub = results.get("subordinate_language", {}).get("count", 0)
total = dom + sub
ratio = {
"dominant_pct": round(dom / total * 100, 1) if total else None,
"subordinate_pct": round(sub / total * 100, 1) if total else None,
"interpretation": (
"More subordinate language" if sub > dom else
"More dominant language" if dom > sub else
"Roughly balanced"
) if total else "No data",
}
return {
"analysis": "extract_dominance_patterns",
"dataset": dataset,
"text_col": text_col,
"subreddit_filter": subreddit,
"total_docs_sampled": total_docs,
"categories": results,
"dominance_ratio": ratio,
"caveat": (
"This analysis counts language patterns in text, not visual image content. "
"It reflects how women are described in text, not how they appear in images. "
"For image-based analysis use analyze_image_sample."
),
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 6. Analysis execution
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _ts_to_date(series: pd.Series) -> pd.Series:
return pd.to_datetime(series, unit="s", utc=True)
def _normalize_filters(
filters: dict | None = None,
subreddit: str | None = None,
date_from: str | None = None,
date_to: str | None = None,
min_score: float | None = None,
) -> dict:
merged = dict(filters or {})
if subreddit:
merged["subreddit"] = subreddit
if date_from:
merged["date_from"] = date_from
if date_to:
merged["date_to"] = date_to
if min_score is not None:
merged["min_score"] = min_score
return merged
def _apply_filters(df: pd.DataFrame, filters: dict | None = None) -> pd.DataFrame:
if not filters:
return df
filtered = df
if filters.get("subreddit") and "subreddit" in filtered.columns:
filtered = filtered[filtered["subreddit"] == filters["subreddit"]]
if filters.get("author") and "author" in filtered.columns:
filtered = filtered[filtered["author"] == filters["author"]]
if filters.get("min_score") is not None and "score" in filtered.columns:
filtered["score"] = pd.to_numeric(filtered["score"], errors="coerce")
filtered = filtered[filtered["score"] >= filters["min_score"]]
if ("date_from" in filters or "date_to" in filters) and "created_utc" in filtered.columns:
created = _ts_to_date(filtered["created_utc"])
if filters.get("date_from"):
filtered = filtered[created >= pd.Timestamp(filters["date_from"], tz="UTC")]
created = _ts_to_date(filtered["created_utc"])
if filters.get("date_to"):
filtered = filtered[created <= pd.Timestamp(filters["date_to"], tz="UTC")]
return filtered
def _save_csv(df: pd.DataFrame, stem: str) -> str:
path = OUTPUTS_DIR / f"{stem}.csv"
df.to_csv(path, index=False)
return str(path)
def _save_fig(fig, stem: str) -> str:
path = OUTPUTS_DIR / f"{stem}.png"
pio.write_image(fig, str(path), scale=2)
return str(path)
def count_by_group(dataset: str, group_col: str, top_n: int = 30, filters: dict | None = None) -> dict:
"""Count rows grouped by a column. Returns sorted table + bar chart."""
filter_cols = [c for c in ["subreddit", "author", "score", "created_utc"] if c != group_col]
df = _load(dataset, columns=list(dict.fromkeys([group_col] + filter_cols)))
df = _apply_filters(df, filters)
counts = (
df.groupby(group_col, dropna=False)
.size().reset_index(name="count")
.sort_values("count", ascending=False).head(top_n)
)
stem = f"count_by_{group_col}_{dataset}"
saved_csv = _save_csv(counts, stem)
fig = px.bar(
counts.sort_values("count"), x="count", y=group_col, orientation="h",
title=f"Count by {group_col}", labels={"count": "Count", group_col: group_col},
)
fig.update_layout(yaxis={"categoryorder": "total ascending"})
try:
saved_png = _save_fig(fig, stem)
except Exception:
saved_png = None
return {
"analysis": "count_by_group", "dataset": dataset, "group_col": group_col,
"filters": filters or {}, "total_rows": len(df),
"table": counts.to_dict(orient="records"),
"saved_csv": saved_csv, "saved_png": saved_png, "plotly_json": fig.to_json(),
}
def trend_over_time(
dataset: str, freq: str = "M", group_col: str | None = None,
top_groups: int = 8, filters: dict | None = None,
) -> dict:
"""Count posts/comments over time, optionally broken out by a grouping column."""
cols = ["created_utc"] + ([group_col] if group_col else []) + ["subreddit", "author", "score"]
cols = list(dict.fromkeys(cols))
df = _load(dataset, columns=cols)
df = _apply_filters(df, filters)
df["period"] = _ts_to_date(df["created_utc"]).dt.to_period(freq).astype(str)
if group_col:
top = df[group_col].value_counts().head(top_groups).index.tolist()
df = df[df[group_col].isin(top)]
counts = (
df.groupby(["period", group_col]).size()
.reset_index(name="count").sort_values("period")
)
fig = px.line(counts, x="period", y="count", color=group_col,
title=f"Activity over time by {group_col}")
else:
counts = (
df.groupby("period").size()
.reset_index(name="count").sort_values("period")
)
fig = px.line(counts, x="period", y="count", title="Activity over time")
stem = f"trend_{dataset}_{group_col or 'all'}_{freq}"
saved_csv = _save_csv(counts, stem)
try:
saved_png = _save_fig(fig, stem)
except Exception:
saved_png = None
return {
"analysis": "trend_over_time", "dataset": dataset, "freq": freq,
"group_col": group_col, "filters": filters or {},
"table": counts.to_dict(orient="records"),
"saved_csv": saved_csv, "saved_png": saved_png, "plotly_json": fig.to_json(),
}
def summary_stats(
dataset: str, value_col: str, group_col: str | None = None,
top_n: int = 30, filters: dict | None = None,
) -> dict:
"""Descriptive statistics for a numeric column, optionally by group."""
cols = [value_col] + ([group_col] if group_col else []) + ["subreddit", "author", "score", "created_utc"]
cols = list(dict.fromkeys(cols))
df = _load(dataset, columns=cols)
df = _apply_filters(df, filters)
df[value_col] = pd.to_numeric(df[value_col], errors="coerce")
if group_col:
stats = (
df.groupby(group_col)[value_col]
.agg(["count", "mean", "median", "std", "min", "max"])
.reset_index().sort_values("mean", ascending=False).head(top_n).round(2)
)
else:
raw = df[value_col].describe().round(2)
stats = raw.reset_index()
stats.columns = ["stat", "value"]
stem = f"stats_{value_col}_{group_col or 'all'}_{dataset}"
saved_csv = _save_csv(stats, stem)
try:
if group_col:
fig = px.bar(stats, x=group_col, y="mean", error_y="std",
title=f"{value_col} by {group_col}",
labels={"mean": f"Mean {value_col}"})
else:
fig = px.histogram(df[value_col].dropna(), nbins=50,
title=f"Distribution of {value_col}",
labels={"value": value_col})
saved_png = _save_fig(fig, stem)
plotly_json = fig.to_json()
except Exception:
saved_png = None
plotly_json = None
return {
"analysis": "summary_stats", "dataset": dataset, "value_col": value_col,
"group_col": group_col, "filters": filters or {},
"n_total": len(df), "n_missing": int(df[value_col].isna().sum()),
"table": stats.to_dict(orient="records"),
"saved_csv": saved_csv, "saved_png": saved_png, "plotly_json": plotly_json,
}
def top_posts(
dataset: str = "posts", n: int = 20,
subreddit: str | None = None, text_col: str = "title",
filters: dict | None = None,
) -> dict:
"""Return the highest-scoring posts, optionally filtered to a subreddit."""
filters = _normalize_filters(filters=filters, subreddit=subreddit)
cols = [c for c in ["subreddit", "author", text_col, "score", "created_utc"] if c]
df = _load(dataset, columns=cols)
df = _apply_filters(df, filters)
top = df.nlargest(n, "score")[cols].copy()
top["date"] = _ts_to_date(top["created_utc"]).dt.strftime("%Y-%m-%d")
top = top.drop(columns=["created_utc"])
stem = f"top_posts_{subreddit or 'all'}_{dataset}"
saved = _save_csv(top, stem)
return {
"analysis": "top_posts", "dataset": dataset,
"subreddit_filter": subreddit, "filters": filters, "n": n,
"table": top.fillna("").to_dict(orient="records"), "saved_csv": saved,
}
def text_search(
dataset: str, query: str, text_col: str = "body",
n: int = 20, case_sensitive: bool = False,
subreddit: str | None = None, filters: dict | None = None,
) -> dict:
"""Search for a string pattern in a text column."""
filters = _normalize_filters(filters=filters, subreddit=subreddit)
cols = [c for c in ["subreddit", "author", text_col, "score", "created_utc"] if c]
df = _load(dataset, columns=cols)
df = _apply_filters(df, filters)
mask = df[text_col].fillna("").str.contains(query, case=case_sensitive, regex=False)
hits = df[mask].nlargest(n, "score").copy()
hits["date"] = _ts_to_date(hits["created_utc"]).dt.strftime("%Y-%m-%d")
hits = hits.drop(columns=["created_utc"])
stem = f"search_{query[:30].replace(' ', '_')}_{dataset}"
saved = _save_csv(hits, stem)
return {
"analysis": "text_search", "dataset": dataset, "query": query,
"text_col": text_col, "filters": filters,
"n_matches": int(mask.sum()), "n_returned": len(hits),
"table": hits.fillna("").to_dict(orient="records"), "saved_csv": saved,
}
def word_freq(
dataset: str = "corpus_clean", text_col: str = "text_cleaned",
top_n: int = 50, subreddit: str | None = None,
min_length: int = 4, filters: dict | None = None,
) -> dict:
"""Count word frequencies in a text column."""
filters = _normalize_filters(filters=filters, subreddit=subreddit)
cols = list(dict.fromkeys([text_col] + (["subreddit"] if subreddit else []) + ["author", "score", "created_utc"]))
df = _load(dataset, columns=cols)
df = _apply_filters(df, filters)
stop = {
"the","and","for","that","with","this","you","are","was","not",
"have","from","they","will","what","been","when","your","more",
"just","about","like","there","were","would","into","than","then",
"some","also","very","only","over","back","can","out","all","but",
"one","had","has","its","which","their","time","our","who","may",
"after","other","these","those","such","each","him","her","his",
"she","how","did","being","now","way","any","too","much","even",
"get","got","got","could","should","make","made","said","still",
"here","because","really","know","think","going","reddit","post",
"comment","deleted","removed",
}
words = (
df[text_col].fillna("").str.lower()
.str.replace(r"[^a-z\s]", " ", regex=True).str.split().explode()
)
words = words[words.str.len() >= min_length]
words = words[~words.isin(stop)]
counts = words.value_counts().head(top_n).reset_index()
counts.columns = ["word", "count"]
stem = f"wordfreq_{text_col}_{subreddit or 'all'}_{dataset}"
saved_csv = _save_csv(counts, stem)
fig = px.bar(
counts.head(30).sort_values("count"), x="count", y="word", orientation="h",
title="Top words by frequency", labels={"count": "Count", "word": "Word"},
)
fig.update_layout(yaxis={"categoryorder": "total ascending"})
try:
saved_png = _save_fig(fig, stem)
except Exception:
saved_png = None
return {
"analysis": "word_freq", "dataset": dataset, "text_col": text_col,
"subreddit_filter": subreddit, "filters": filters, "total_docs": len(df),
"table": counts.to_dict(orient="records"),
"saved_csv": saved_csv, "saved_png": saved_png, "plotly_json": fig.to_json(),
}
def compare_groups(
dataset: str, group_col: str, value_col: str,
groups: list[str] | None = None, filters: dict | None = None,
) -> dict:
"""Compare a numeric value across groups with descriptive stats."""
cols = list(dict.fromkeys([group_col, value_col, "subreddit", "author", "score", "created_utc"]))
df = _load(dataset, columns=cols)
df = _apply_filters(df, filters)
df[value_col] = pd.to_numeric(df[value_col], errors="coerce")
if groups:
df = df[df[group_col].isin(groups)]
stats = (
df.groupby(group_col)[value_col]
.agg(count="count", mean="mean", median="median", std="std")
.reset_index().sort_values("median", ascending=False).round(3)
)
stem = f"compare_{group_col}_{value_col}_{dataset}"
saved_csv = _save_csv(stats, stem)
fig = px.bar(stats, x=group_col, y="median", error_y="std",
title=f"{value_col} by {group_col} (median Β± std)",
labels={"median": f"Median {value_col}"})
try:
saved_png = _save_fig(fig, stem)
except Exception:
saved_png = None
return {
"analysis": "compare_groups", "dataset": dataset,
"group_col": group_col, "value_col": value_col,
"filters": filters or {}, "groups_compared": stats[group_col].tolist(),
"table": stats.to_dict(orient="records"),
"saved_csv": saved_csv, "saved_png": saved_png, "plotly_json": fig.to_json(),
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 7. Core agent loop
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
MODEL = "claude-opus-4-6"
TOOLS = [
{
"name": "list_datasets",
"description": (
"List cached dataset metadata: paths, row counts, columns, subreddits, and date ranges. "
"Use this to inspect the available data without loading full tables."
),
"input_schema": {
"type": "object",
"properties": {"refresh": {"type": "boolean", "default": False,
"description": "Recompute metadata from source parquets instead of using the cache."}},
"required": [],
},
},
{
"name": "sample_rows",
"description": "Return a small deterministic preview of rows from a dataset, optionally filtered and column-limited.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean", "titles"]},
"n": {"type": "integer", "default": 5},
"filters": {"type": "object", "description": "Optional equality filters, e.g. {\"subreddit\": \"GOONED\"}"},
"columns": {"type": "array", "items": {"type": "string"}, "description": "Optional subset of columns to preview."},
},
"required": ["dataset"],
},
},
{
"name": "count_by_group",
"description": "Count rows in a dataset grouped by one column, with optional shared filters.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean", "titles"]},
"group_col": {"type": "string"},
"top_n": {"type": "integer", "default": 30},
"filters": {"type": "object"},
},
"required": ["dataset", "group_col"],
},
},
{
"name": "trend_over_time",
"description": "Count rows over time, optionally split by one grouping column, with optional shared filters.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean", "titles"]},
"freq": {"type": "string", "enum": ["D", "W", "M", "Q", "Y"], "default": "M"},
"group_col": {"type": "string"},
"top_groups": {"type": "integer", "default": 8},
"filters": {"type": "object"},
},
"required": ["dataset"],
},
},
{
"name": "summary_stats",
"description": "Descriptive statistics for a numeric column, optionally grouped and filtered.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean", "titles"]},
"value_col": {"type": "string"},
"group_col": {"type": "string"},
"top_n": {"type": "integer", "default": 30},
"filters": {"type": "object"},
},
"required": ["dataset", "value_col"],
},
},
{
"name": "top_posts",
"description": "Return the highest-scoring posts, optionally filtered by subreddit or shared filters.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "titles"], "default": "posts"},
"n": {"type": "integer", "default": 20},
"subreddit": {"type": "string"},
"text_col": {"type": "string", "default": "title"},
"filters": {"type": "object"},
},
"required": [],
},
},
{
"name": "text_search",
"description": "Search for a phrase in a text column and return top matching rows.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean", "titles"]},
"query": {"type": "string"},
"text_col": {"type": "string", "default": "body"},
"n": {"type": "integer", "default": 20},
"subreddit": {"type": "string"},
"filters": {"type": "object"},
},
"required": ["dataset", "query"],
},
},
{
"name": "word_freq",
"description": "Count word frequencies in a text column with optional shared filters.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean", "titles"], "default": "corpus_clean"},
"text_col": {"type": "string", "default": "text_cleaned"},
"top_n": {"type": "integer", "default": 50},
"subreddit": {"type": "string"},
"min_length": {"type": "integer", "default": 4},
"filters": {"type": "object"},
},
"required": [],
},
},
{
"name": "compare_groups",
"description": "Compare one numeric column across groups with optional shared filters.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean", "titles"]},
"group_col": {"type": "string"},
"value_col": {"type": "string"},
"groups": {"type": "array", "items": {"type": "string"}},
"filters": {"type": "object"},
},
"required": ["dataset", "group_col", "value_col"],
},
},
{
"name": "extract_frequency_patterns",
"description": "Mine text for frequency and duration language across the full dataset.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean"], "default": "comments"},
"text_col": {"type": "string", "default": "body"},
"subreddit": {"type": "string"},
"n_examples": {"type": "integer", "default": 5},
"sample_size": {"type": "integer", "default": 5000000},
},
"required": [],
},
},
{
"name": "extract_dominance_patterns",
"description": "Count dominant vs subordinate language in text, not images.",
"input_schema": {
"type": "object",
"properties": {
"dataset": {"type": "string", "enum": ["posts", "comments", "corpus_clean"], "default": "comments"},
"text_col": {"type": "string", "default": "body"},
"subreddit": {"type": "string"},
"sample_size": {"type": "integer", "default": 5000000},
},
"required": [],
},
},
{
"name": "analyze_image_sample",
"description": "Run vision coding on a sample of image posts using Qwen2-VL via Together AI (no content filters). Always provide a coding_scheme for research use.",
"input_schema": {
"type": "object",
"properties": {
"question": {"type": "string"},
"subreddit": {"type": "string"},
"n_sample": {"type": "integer", "default": 100, "description": "Number of images to code. No hard cap β set to 500+ for large analyses."},
"coding_scheme": {"type": "object", "description": "Dict of {label: definition}. Always provide this for research questions."},
},
"required": ["question"],
},
},
{
"name": "export_reliability_sample",
"description": "Export a stratified random sample of coded images for human validation. Run after analyze_image_sample.",
"input_schema": {
"type": "object",
"properties": {
"source_csv": {"type": "string", "description": "Path to image_analysis CSV. Defaults to most recent."},
"n": {"type": "integer", "default": 200},
"random_state": {"type": "integer", "default": 42},
},
"required": [],
},
},
{
"name": "compute_reliability",
"description": "Compute Cohen's kappa between model and human codes after the human_label column has been filled in.",
"input_schema": {
"type": "object",
"properties": {
"human_csv_path": {"type": "string", "description": "Path to completed reliability_sample.csv. Defaults to outputs/reliability_sample.csv."},
},
"required": [],
},
},
]
TOOL_FN_MAP = {
"list_datasets": lambda args: list_datasets(**args),
"sample_rows": lambda args: sample_rows(**args),
"count_by_group": lambda args: count_by_group(**args),
"trend_over_time": lambda args: trend_over_time(**args),
"summary_stats": lambda args: summary_stats(**args),
"top_posts": lambda args: top_posts(**args),
"text_search": lambda args: text_search(**args),
"word_freq": lambda args: word_freq(**args),
"compare_groups": lambda args: compare_groups(**args),
"extract_frequency_patterns": lambda args: extract_frequency_patterns(**args),
"extract_dominance_patterns": lambda args: extract_dominance_patterns(**args),
"analyze_image_sample": lambda args: analyze_image_sample(**args),
"export_reliability_sample": lambda args: export_reliability_sample(**args),
"compute_reliability": lambda args: compute_reliability(**args),
}
def _safe_str(obj: object) -> object:
"""Recursively encode any non-ASCII strings as JSON-safe escaped text."""
if isinstance(obj, str):
return obj.encode("ascii", errors="backslashreplace").decode("ascii")
if isinstance(obj, dict):
return {k: _safe_str(v) for k, v in obj.items()}
if isinstance(obj, list):
return [_safe_str(item) for item in obj]
return obj
def _compact_result(result: object) -> dict:
if not isinstance(result, dict):
return {"value": result}
compact = {}
for key in ("analysis", "dataset", "group_col", "value_col", "query", "filters",
"n_matches", "n_returned", "n_total", "groups_compared", "saved_csv", "saved_png", "error"):
if key in result and result.get(key) is not None:
compact[key] = result[key]
table = result.get("table")
if isinstance(table, list):
compact["table_preview"] = table[:3]
compact["table_rows"] = len(table)
return compact
def _conversation_state_summary(turns: list[dict] | None) -> str:
if not turns:
return "No prior analytical state."
summary = []
for idx, turn in enumerate(turns[-3:], start=1):
summary.append({
"turn": idx,
"question": _safe_str(turn.get("question", "")),
"answer": _safe_str(turn.get("answer", "")),
"tool_calls": [
{"tool": tc.get("tool"), "args": _safe_str(tc.get("args", {})),
"result": _safe_str(_compact_result(tc.get("result")))}
for tc in turn.get("tool_calls", [])
],
"artifacts": turn.get("artifacts", []),
})
return json.dumps(summary, default=str, indent=2)
def _tool_names(tools: list[dict]) -> list[str]:
return [t["name"] for t in tools]
def _tool_subset(allowed_tools: list[str]) -> list[dict]:
allowed = set(allowed_tools)
return [t for t in TOOLS if t["name"] in allowed]
def _system_prompt(route_mode: str, route_guidance: str, conversation_state: str) -> str:
metadata = get_dataset_metadata()
dataset_lines = []
for name, info in metadata.items():
if not info.get("available"):
continue
date_range = info.get("date_range") or {}
dataset_lines.append(
f"- {name}: {info.get('rows')} rows; columns={list(info.get('columns', {}).keys())}; "
f"date_range={date_range or 'n/a'}"
)
dataset_summary = "\n".join(dataset_lines)
return f"""You are a question-driven data analysis agent working over local Reddit datasets.
Available dataset metadata:
{dataset_summary}
Current route mode: {route_mode}
Route guidance: {route_guidance}
Prior analytical state:
{conversation_state}
Rules:
1. Use the route guidance and only the provided tools.
2. Inspect metadata or row previews before making assumptions when the schema is unclear.
3. Run actual tools for numbers; do not guess.
4. Prefer one minimal reproducible tool path over exploratory tool spam.
5. Distinguish direct findings from caveats.
6. If prior turns already produced a relevant result, reuse that context instead of recomputing unless the user asks for a change.
7. Answer with this structure: direct answer, what was analysed, method, caveats.
8. ALWAYS prefer tools that produce charts (trend_over_time, count_by_group, compare_groups, summary_stats, word_freq) over plain text summaries when the question is quantitative. Every numeric answer should have a chart.
9. For questions about images or visual content, use analyze_image_sample. It reads from raw CSV files with image URLs β no separate setup needed. ALWAYS generate an explicit coding_scheme dict (with label names as keys and definitions as values) before calling this tool β never leave coding_scheme null for a research question.
10. After a large image coding run, offer to run export_reliability_sample to generate a human validation set, then compute_reliability once the user has filled in the human_label column.
11. The dataset covers 30 subreddits including GOONED, GOONEDISBACK, GoonCaves, girlgooners, and more. Use subreddit filters to drill into specific communities."""
def run_agent(
question: str,
history: list[dict] | None = None,
turns: list[dict] | None = None,
analysis_context: list[dict] | None = None,
conversation_state: list[dict] | None = None,
) -> dict:
"""Run the agent for a user question with deterministic routing and structured prior state."""
try:
return _run_agent_inner(question, history, turns, analysis_context, conversation_state)
except UnicodeEncodeError as exc:
tb = _traceback.format_exc()
raise RuntimeError(
f"Unicode encoding error (non-ASCII character in data pipeline).\n\n"
f"Detail: {exc}\n\nTraceback:\n{tb}"
) from exc
def _run_agent_inner(
question: str,
history: list[dict] | None = None,
turns: list[dict] | None = None,
analysis_context: list[dict] | None = None,
conversation_state: list[dict] | None = None,
) -> dict:
client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
prior_turns = turns or analysis_context or conversation_state or []
route = route_question(question)
available_tools = _tool_subset(route.allowed_tools)
safe_history = [_safe_str(msg) for msg in (history or [])]
messages = safe_history
messages.append({"role": "user", "content": _safe_str(question)})
tool_calls_log = []
plotly_jsons = []
total_input_tokens = 0
total_output_tokens = 0
system = _system_prompt(
route_mode=route.mode,
route_guidance=route.guidance,
conversation_state=_conversation_state_summary(prior_turns),
)
while True:
safe_messages = _safe_str(messages)
safe_system = _safe_str(system)
try:
response = client.messages.create(
model=MODEL,
max_tokens=4096,
system=safe_system,
tools=available_tools,
messages=safe_messages,
)
except UnicodeEncodeError:
stripped_messages = json.loads(json.dumps(safe_messages, default=str, ensure_ascii=True))
stripped_system = safe_system.encode("ascii", errors="ignore").decode("ascii")
response = client.messages.create(
model=MODEL,
max_tokens=4096,
system=stripped_system,
tools=available_tools,
messages=stripped_messages,
)
text_parts = [block.text for block in response.content if block.type == "text"]
tool_use_blocks = [block for block in response.content if block.type == "tool_use"]
total_input_tokens += getattr(response.usage, "input_tokens", 0) or 0
total_output_tokens += getattr(response.usage, "output_tokens", 0) or 0
if response.stop_reason == "end_turn" or not tool_use_blocks:
# Claude Opus 4.6 pricing: $15/M input, $75/M output
cost_usd = (total_input_tokens / 1_000_000 * 15.0) + (total_output_tokens / 1_000_000 * 75.0)
return {
"answer": "\n".join(text_parts).strip(),
"tool_calls": tool_calls_log,
"plotly_json": plotly_jsons[-1] if plotly_jsons else None,
"plotly_jsons": plotly_jsons,
"route": route.mode,
"allowed_tools": _tool_names(available_tools),
"usage": {
"input_tokens": total_input_tokens,
"output_tokens": total_output_tokens,
"cost_usd": round(cost_usd, 4),
},
}
tool_results = []
for block in tool_use_blocks:
fn = TOOL_FN_MAP.get(block.name)
if fn is None:
result = {"error": f"Unknown tool: {block.name}"}
else:
try:
result = fn(block.input)
if isinstance(result, dict) and result.get("plotly_json"):
plotly_jsons.append(result["plotly_json"])
except Exception as exc:
result = {"error": str(exc)}
safe_result = _safe_str(result)
tool_calls_log.append({"tool": block.name, "args": block.input, "result": result})
tool_results.append({
"type": "tool_result",
"tool_use_id": block.id,
"content": json.dumps(safe_result, default=str, ensure_ascii=True),
})
assistant_content = [_safe_str(block.model_dump()) for block in response.content]
messages.append({"role": "assistant", "content": assistant_content})
messages.append({"role": "user", "content": tool_results})
|