| 1 | """X corpus judging for retrieve-judge-retry. |
| 2 | |
| 3 | No I/O: judges items already retrieved. Reuses relevance.token_overlap_relevance. |
| 4 | |
| 5 | The judge detects off-topic floods and determines which extracted handles should |
| 6 | be promoted to the FROM lane based on on-topic post ratio, not frequency. |
| 7 | """ |
| 8 | |
| 9 | from collections import Counter |
| 10 | from typing import Any, Dict, List, Optional, Set, Tuple |
| 11 | |
| 12 | from . import relevance |
| 13 | |
| 14 | # Minimum on-topic ratio for the overall corpus. Below this, the engine should |
| 15 | # retry with a wider keyword query. Rome measured ~0.2 (8/40 on-topic). |
| 16 | CORPUS_ON_TOPIC_FLOOR = 0.4 |
| 17 | |
| 18 | # Minimum on-topic ratio for a handle's posts to qualify for FROM promotion. |
| 19 | HANDLE_ON_TOPIC_FLOOR = 0.5 |
| 20 | |
| 21 | # Minimum on-topic keyword hits before a handle can be promoted to FROM lane. |
| 22 | # Prevents promoting handles that appeared in thin phrase hits with only 1 match. |
| 23 | MIN_ON_TOPIC_HITS = 2 |
| 24 | |
| 25 | |
| 26 | # Ambiguous short tokens that require case-sensitive matching to avoid |
| 27 | # pronoun/acronym collisions. E.g., "US" (country) vs "us" (pronoun). |
| 28 | # For these tokens, require the text to contain the uppercase form (acronym) |
| 29 | # rather than just the lowercase form (common word). |
| 30 | _CASE_SENSITIVE_ACRONYMS = frozenset({'us'}) |
| 31 | |
| 32 | |
| 33 | def _compute_relevance(query: str, text: str) -> float: |
| 34 | """Compute relevance score for a post against the topic query. |
| 35 | |
| 36 | Returns 0.0 for empty/stopword-only queries to avoid treating all items |
| 37 | as equally relevant (the shared relevance module returns 0.5 for empty |
| 38 | queries as a neutral fallback, but x_judge needs strict filtering). |
| 39 | |
| 40 | Uses case-sensitive matching for ambiguous short tokens like 'us' to |
| 41 | distinguish country acronym 'US' from pronoun 'us'. |
| 42 | """ |
| 43 | import re |
| 44 | if not query or not text: |
| 45 | return 0.0 |
| 46 | |
| 47 | q_tokens = relevance.tokenize(query) |
| 48 | if not q_tokens: |
| 49 | return 0.0 # All query tokens were stopwords |
| 50 | |
| 51 | # Check for ambiguous acronyms that need case-sensitive handling |
| 52 | t_tokens = relevance.tokenize(text) |
| 53 | filtered_q_tokens = set(q_tokens) |
| 54 | for acronym in _CASE_SENSITIVE_ACRONYMS: |
| 55 | if acronym in q_tokens and acronym in t_tokens: |
| 56 | # Text has the token, but we need to check if it's the acronym (US) |
| 57 | # or the common word (us). If text only has lowercase, don't count it. |
| 58 | has_uppercase = bool(re.search(rf'\b{acronym.upper()}\b', text)) |
| 59 | has_lowercase = bool(re.search(rf'\b{acronym}\b', text)) |
| 60 | if has_lowercase and not has_uppercase: |
| 61 | # Text only has lowercase version (pronoun) - don't count this |
| 62 | # token in query overlap. This effectively removes "us" from |
| 63 | # contributing to the score when text has only the pronoun. |
| 64 | t_tokens = t_tokens - {acronym} |
| 65 | # Also remove from query for this calculation to avoid |
| 66 | # penalizing the overall coverage ratio |
| 67 | filtered_q_tokens = filtered_q_tokens - {acronym} |
| 68 | |
| 69 | # If filtering removed all query tokens, fall back to 0 |
| 70 | if not filtered_q_tokens: |
| 71 | return 0.0 |
| 72 | |
| 73 | overlap_tokens = filtered_q_tokens & t_tokens |
| 74 | if not overlap_tokens: |
| 75 | return 0.0 |
| 76 | |
| 77 | # Compute simplified relevance: coverage ratio |
| 78 | # This is a simpler version of token_overlap_relevance that uses |
| 79 | # the filtered tokens rather than re-tokenizing the original text |
| 80 | coverage = len(overlap_tokens) / len(filtered_q_tokens) |
| 81 | return coverage |
| 82 | |
| 83 | |
| 84 | def judge_x_corpus( |
| 85 | items: List[Dict[str, Any]], |
| 86 | topic: str, |
| 87 | *, |
| 88 | ranking_query: str = "", |
| 89 | ) -> Dict[str, Any]: |
| 90 | """Judge the retrieved X corpus for on-topic ratio. |
| 91 | |
| 92 | Args: |
| 93 | items: List of X items with 'author_handle' and 'text' fields |
| 94 | topic: The user topic (e.g., "Rome") |
| 95 | ranking_query: Optional ranking query for better relevance scoring |
| 96 | |
| 97 | Returns: |
| 98 | Dict with: |
| 99 | - on_topic_ratio: float, fraction of posts that are on-topic |
| 100 | - is_off_topic_flood: bool, True if corpus fails the on-topic floor |
| 101 | - on_topic_items: list, items that passed relevance check |
| 102 | - off_topic_items: list, items that failed relevance check |
| 103 | - handle_stats: dict, per-handle on-topic counts and totals |
| 104 | """ |
| 105 | if not items: |
| 106 | return { |
| 107 | "on_topic_ratio": 1.0, |
| 108 | "is_off_topic_flood": False, |
| 109 | "on_topic_items": [], |
| 110 | "off_topic_items": [], |
| 111 | "handle_stats": {}, |
| 112 | } |
| 113 | |
| 114 | # Use ranking_query if provided, otherwise topic |
| 115 | query = ranking_query or topic |
| 116 | |
| 117 | on_topic_items = [] |
| 118 | off_topic_items = [] |
| 119 | handle_stats: Dict[str, Dict[str, int]] = {} |
| 120 | |
| 121 | for item in items: |
| 122 | handle = (item.get("author_handle") or "").lower() |
| 123 | text = item.get("text") or "" |
| 124 | score = _compute_relevance(query, text) |
| 125 | |
| 126 | # On-topic threshold: relevance.RELEVANCE_FLOOR is 0.1 |
| 127 | is_on_topic = score >= relevance.RELEVANCE_FLOOR |
| 128 | |
| 129 | if is_on_topic: |
| 130 | on_topic_items.append(item) |
| 131 | else: |
| 132 | off_topic_items.append(item) |
| 133 | |
| 134 | if handle: |
| 135 | if handle not in handle_stats: |
| 136 | handle_stats[handle] = {"on_topic": 0, "total": 0} |
| 137 | handle_stats[handle]["total"] += 1 |
| 138 | if is_on_topic: |
| 139 | handle_stats[handle]["on_topic"] += 1 |
| 140 | |
| 141 | on_topic_ratio = len(on_topic_items) / len(items) if items else 1.0 |
| 142 | |
| 143 | # Check if top-3 frequency authors have poor on-topic ratio |
| 144 | top_handles = sorted( |
| 145 | handle_stats.items(), |
| 146 | key=lambda x: x[1]["total"], |
| 147 | reverse=True, |
| 148 | )[:3] |
| 149 | top_authors_off_topic = all( |
| 150 | stats["on_topic"] / stats["total"] < HANDLE_ON_TOPIC_FLOOR |
| 151 | for _, stats in top_handles |
| 152 | if stats["total"] > 0 |
| 153 | ) if top_handles else False |
| 154 | |
| 155 | is_off_topic_flood = ( |
| 156 | on_topic_ratio < CORPUS_ON_TOPIC_FLOOR |
| 157 | or (top_authors_off_topic and len(on_topic_items) < MIN_ON_TOPIC_HITS) |
| 158 | ) |
| 159 | |
| 160 | return { |
| 161 | "on_topic_ratio": on_topic_ratio, |
| 162 | "is_off_topic_flood": is_off_topic_flood, |
| 163 | "on_topic_items": on_topic_items, |
| 164 | "off_topic_items": off_topic_items, |
| 165 | "handle_stats": handle_stats, |
| 166 | } |
| 167 | |
| 168 | |
| 169 | def promotable_handles( |
| 170 | items: List[Dict[str, Any]], |
| 171 | topic: str, |
| 172 | extracted_handles: List[str], |
| 173 | *, |
| 174 | explicit_handles: Optional[List[str]] = None, |
| 175 | ranking_query: str = "", |
| 176 | ) -> Tuple[List[str], List[str]]: |
| 177 | """Determine which handles should be promoted to the FROM lane. |
| 178 | |
| 179 | Split FROM promotion: |
| 180 | - Explicit handles (--x-handle/--x-related): always promoted, no AND topic |
| 181 | - Extracted handles: promoted only if: |
| 182 | - ≥MIN_ON_TOPIC_HITS on-topic keyword hits AND |
| 183 | - author on-topic ratio ≥ HANDLE_ON_TOPIC_FLOOR |
| 184 | - These pulls AND the topic (from:handle Rome) |
| 185 | |
| 186 | Args: |
| 187 | items: List of X items with 'author_handle' and 'text' fields |
| 188 | topic: The user topic |
| 189 | extracted_handles: Handles extracted from entity_extract |
| 190 | explicit_handles: Explicit --x-handle/--x-related handles |
| 191 | ranking_query: Optional ranking query for relevance scoring |
| 192 | |
| 193 | Returns: |
| 194 | Tuple of (explicit_promotable, extracted_promotable): |
| 195 | - explicit_promotable: handles that get FROM without AND topic |
| 196 | - extracted_promotable: handles that get FROM with AND topic |
| 197 | """ |
| 198 | explicit_set = { |
| 199 | h.lower().lstrip("@") |
| 200 | for h in (explicit_handles or []) |
| 201 | if h and h.strip() |
| 202 | } |
| 203 | |
| 204 | # Judge corpus to get handle stats |
| 205 | judgment = judge_x_corpus(items, topic, ranking_query=ranking_query) |
| 206 | handle_stats = judgment["handle_stats"] |
| 207 | |
| 208 | explicit_promotable = [] |
| 209 | extracted_promotable = [] |
| 210 | |
| 211 | for handle in extracted_handles: |
| 212 | handle_lower = handle.lower().lstrip("@") |
| 213 | |
| 214 | # Explicit handles always promoted (no AND topic) |
| 215 | if handle_lower in explicit_set: |
| 216 | explicit_promotable.append(handle) |
| 217 | continue |
| 218 | |
| 219 | # Check if handle qualifies for extracted promotion |
| 220 | stats = handle_stats.get(handle_lower, {"on_topic": 0, "total": 0}) |
| 221 | |
| 222 | # Need ≥MIN_ON_TOPIC_HITS on-topic posts |
| 223 | if stats["on_topic"] < MIN_ON_TOPIC_HITS: |
| 224 | continue |
| 225 | |
| 226 | # Need ≥HANDLE_ON_TOPIC_FLOOR ratio |
| 227 | if stats["total"] > 0: |
| 228 | ratio = stats["on_topic"] / stats["total"] |
| 229 | if ratio >= HANDLE_ON_TOPIC_FLOOR: |
| 230 | extracted_promotable.append(handle) |
| 231 | |
| 232 | # Also check explicit handles not in extracted list |
| 233 | for handle in (explicit_handles or []): |
| 234 | handle_lower = handle.lower().lstrip("@") |
| 235 | if handle_lower not in [h.lower() for h in explicit_promotable]: |
| 236 | if handle_lower not in [h.lower() for h in extracted_handles]: |
| 237 | explicit_promotable.append(handle) |
| 238 | |
| 239 | return explicit_promotable, extracted_promotable |
| 240 | |
| 241 | |
| 242 | def should_retry_x_search( |
| 243 | items: List[Dict[str, Any]], |
| 244 | topic: str, |
| 245 | *, |
| 246 | ranking_query: str = "", |
| 247 | depth: str = "default", |
| 248 | ) -> bool: |
| 249 | """Determine if X search should retry with wider keyword query. |
| 250 | |
| 251 | Skip retry on quick/mock (same as Phase 2). |
| 252 | |
| 253 | Args: |
| 254 | items: Retrieved X items |
| 255 | topic: The user topic |
| 256 | ranking_query: Optional ranking query |
| 257 | depth: Search depth ("quick", "default", "deep") |
| 258 | |
| 259 | Returns: |
| 260 | True if retry is warranted |
| 261 | """ |
| 262 | if depth == "quick": |
| 263 | return False |
| 264 | |
| 265 | if not items: |
| 266 | return False # Nothing to judge, no retry |
| 267 | |
| 268 | judgment = judge_x_corpus(items, topic, ranking_query=ranking_query) |
| 269 | return judgment["is_off_topic_flood"] |
| 270 | |
| 271 | |
| 272 | def prune_off_topic_items( |
| 273 | items: List[Dict[str, Any]], |
| 274 | topic: str, |
| 275 | *, |
| 276 | ranking_query: str = "", |
| 277 | ) -> List[Dict[str, Any]]: |
| 278 | """Prune off-topic items before the pool. |
| 279 | |
| 280 | Eight on-topic items with 32 pruned → ok with 8. |
| 281 | Zero on-topic → no-results, not ok with 40 junk. |
| 282 | |
| 283 | Args: |
| 284 | items: Retrieved X items |
| 285 | topic: The user topic |
| 286 | ranking_query: Optional ranking query |
| 287 | |
| 288 | Returns: |
| 289 | Only on-topic items |
| 290 | """ |
| 291 | judgment = judge_x_corpus(items, topic, ranking_query=ranking_query) |
| 292 | return judgment["on_topic_items"] |
| 293 |