File size: 35,350 Bytes
1d9bd9b | 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 | """Moonley agentic TOOL REGISTRY. Loads the corpus once (lean β no BM25) and exposes the tools the
ReAct controller can call. Each tool returns a list of compact case dicts (doc_id + the fields the
LLM needs to reason) or a small structured result. Most tools are ports of serve.py primitives;
the statute tools are new (see themis-statute-layer). keyword_search(BM25) is intentionally omitted
here β it is 68s/query on the Mac; it runs on the GPU box in production.
Usage: from tools import Corpus ; C = Corpus(DATA, STATUTE_DIR) ; C.vector_search("...", k=8)
"""
import json, os, re, difflib
import numpy as np
from collections import Counter, defaultdict
from sentence_transformers import SentenceTransformer, CrossEncoder
from case_summary import case_summary_record, load_case_summaries
from statute_crosswalk import load_default_crosswalk
from statute_library import ExactStatuteLibrary
BGE_Q = "Represent this sentence for searching relevant passages: "
_NAME_STOP = {"v","vs","of","and","the","ors","anr","etc","state","union","govt","government","in","re",
"others","another","ltd","co","pvt","dead","thr","lrs","alias","through","etc","anrs","ms",
"shri","smt","sri","mr","mrs","dr","justice","sh","kum","mohd"}
BAD_STATUS = {"overruled", "per_incuriam", "doubted"}
def _clean(s):
if not s: return ""
return re.sub(r"\s+", " ", s).strip()
def _ntok(s):
"""Tokenize a case name, MERGING runs of single letters so abbreviations match: U.P.->up, A.K.->ak."""
out = []; buf = ""
for t in re.findall(r"[a-z]+", (s or "").lower()):
if len(t) == 1: buf += t
else:
if buf: out.append(buf); buf = ""
out.append(t)
if buf: out.append(buf)
return out
class Corpus:
def __init__(self, data_dir, statute_dir, device="cpu"):
self.device = device
print("[tools] loading corpus ...", flush=True)
self.texts = []; self.chunk_doc = []
with open(os.path.join(data_dir, "escr_chunks.jsonl"), encoding="utf-8") as f:
for l in f:
c = json.loads(l); self.texts.append(c["text"]); self.chunk_doc.append(c["doc_id"])
self.M = np.load(os.path.join(data_dir, "escr_vectors.npy"), mmap_mode="r") # mmap: ~2GB off resident RSS (pages fault in on the M@qv scan)
self.chunk_doc_arr = np.array(self.chunk_doc)
self.doc_chunks = defaultdict(list)
for i, d in enumerate(self.chunk_doc): self.doc_chunks[d].append(i)
self.meta = {}
for l in open(os.path.join(data_dir, "escr_meta.jsonl"), encoding="utf-8"):
m = json.loads(l); self.meta[m["doc_id"]] = m
# A judgment is eligible for research only when both its metadata and
# source-derived text are present in the active release. Metadata-only
# rows and graph stubs may remain useful for audit work, but they must
# never become search results, recommendations, or chat evidence.
self.eligible_doc_ids = {
d for d, cis in self.doc_chunks.items()
if d in self.meta and any(_clean(self.texts[i]) for i in cis)
}
self.goodlaw = {}
# good_law_v2 (classify_treatments.py rollup β real statuses) wins over the legacy file
_gl = "good_law_v2.jsonl" if os.path.exists(os.path.join(data_dir, "good_law_v2.jsonl")) else "good_law.jsonl"
for l in open(os.path.join(data_dir, _gl), encoding="utf-8"):
g = json.loads(l); self.goodlaw[g["doc_id"]] = g
# identity ledger (build_ledger.py): decision-year fix, bench ints, sibling canonicals
self.decision_year = {}; self.bench_n = {}; self.canonical = set(); self.cluster_of = {}
_lp = os.path.join(data_dir, "corpus_ledger.jsonl")
if os.path.exists(_lp):
for l in open(_lp, encoding="utf-8"):
r = json.loads(l); d = r["doc_id"]
if r.get("decision_year"): self.decision_year[d] = r["decision_year"]
self.bench_n[d] = r.get("bench_n", 0)
if r.get("canonical", True): self.canonical.add(d)
if r.get("cluster_id"): self.cluster_of[d] = r["cluster_id"]
print(f"[tools] ledger: {len(self.decision_year)} decision-years, "
f"{len(self.cluster_of)} sibling-clustered docs", flush=True)
# famous-name aliases mined from citing sentences (build_citation_graph.py)
self.aliases = {}
_ap = os.path.join(data_dir, "case_aliases.json")
if os.path.exists(_ap):
self.aliases = {k.lower(): v for k, v in json.load(open(_ap, encoding="utf-8")).items()}
print(f"[tools] aliases: {len(self.aliases)}", flush=True)
self.in_edges = defaultdict(list); self.out_edges = defaultdict(list); self.edge_meta = {}
self.cite_indeg = Counter()
# edges_v2 (body-text parse, ~25x the legacy graph) wins over the headnote-only file
_ep = "edges_v2.jsonl" if os.path.exists(os.path.join(data_dir, "edges_v2.jsonl")) else "edges.jsonl"
for l in open(os.path.join(data_dir, _ep), encoding="utf-8"):
e = json.loads(l); f, t = e["from"], e["target"]
self.out_edges[f].append(t); self.in_edges[t].append(f)
self.edge_meta[(f, t)] = {"treatment": e.get("treatment"), "method": e.get("method")}
if e.get("method") in ("cite", "body", "headnote"): self.cite_indeg[t] += 1
if _ep == "edges_v2.jsonl":
print(f"[tools] edges_v2: {sum(len(v) for v in self.out_edges.values())} edges", flush=True)
# treatment overrides from the classifier (finer than the rollup)
_tp = os.path.join(data_dir, "edges_treatment.jsonl")
if os.path.exists(_tp):
n = 0
for l in open(_tp, encoding="utf-8"):
r = json.loads(l); k = (r["from"], r["target"])
if k in self.edge_meta: self.edge_meta[k]["treatment"] = r["treatment"]; n += 1
print(f"[tools] treatments: {n} classified edges", flush=True)
# synthetic headnotes (backfill_headnotes.py) β fill the 70s-80s crater for skim/judge/view
self.syn_held = {}
_sp = os.path.join(data_dir, "synthetic_headnotes.jsonl")
if os.path.exists(_sp):
for l in open(_sp, encoding="utf-8"):
r = json.loads(l)
if r.get("held"): self.syn_held[r["doc_id"]] = r
print(f"[tools] synthetic headnotes: {len(self.syn_held)}", flush=True)
# Extraction-time case summaries. This sidecar is the stable hand-off for the future
# Indian Kanoon re-extraction; reporter/synthetic headnotes remain honest interim fallbacks.
self.case_summaries, _summary_file = load_case_summaries(data_dir)
if _summary_file:
print(f"[tools] case summaries: {len(self.case_summaries)} from {_summary_file}", flush=True)
self.name_vocab = set(); self.name_postings = defaultdict(set) # token -> doc_ids (fast name lookup)
self.nc2doc = {}; self.cite_resolver = {} # exact citation lookup (known-item route)
for d, m in self.meta.items():
for w in _ntok(m.get("case_name") or ""): # _ntok merges U.P.->up so abbreviations match
if len(w) >= 4: self.name_vocab.add(w)
if len(w) > 1: self.name_postings[w].add(d)
if m.get("neutral_citation"): self.nc2doc[m["neutral_citation"]] = d
for k in [m.get("neutral_citation")] + (m.get("equivalent_citations") or []):
if k: self.cite_resolver.setdefault(re.sub(r"\s+", " ", k.replace(".", "")).strip().upper(), d)
# statute layer
self.statute_idx = json.load(open(os.path.join(statute_dir, "statute_index.json")))
self.statute_V = np.load(os.path.join(statute_dir, "statute_vectors.npy"))
_statute_records = json.load(open(os.path.join(statute_dir, "all_statutes.json")))
self.statute_texts = [s.get("retrieval_text", "") for s in _statute_records]
self.statute_library = ExactStatuteLibrary.from_env(
fallback_path=os.path.join(statute_dir, "all_statutes.json")
)
self.concord = json.load(open(os.path.join(statute_dir, "concordance.json")))
self.crosswalk = load_default_crosswalk(os.environ.get("THEMIS_SECTION_CROSSWALK", "").strip() or None)
# doc-level HELD-headnote vectors (optional; the $0 representation arm β cleaner signal than OCR chunks)
self.held_V = None
hv = os.path.join(data_dir, "held_vectors.npy")
if os.path.exists(hv):
self.held_V = np.load(hv)
self.held_docs = json.load(open(os.path.join(data_dir, "held_docids.json")))
print(f"[tools] HELD vectors: {self.held_V.shape[0]}", flush=True)
# citation-context vectors (optional; how LATER courts describe each precedent β clean,
# modern, doctrine-level; exists precisely for old landmarks whose own text is OCR-noisy)
self.citectx_V = None
cv = os.path.join(data_dir, "citectx_vectors.npy")
if os.path.exists(cv):
self.citectx_V = np.load(cv, mmap_mode="r")
self.citectx_docs = json.load(open(os.path.join(data_dir, "citectx_docids.json")))
print(f"[tools] CITECTX vectors: {self.citectx_V.shape[0]}", flush=True)
self.st = SentenceTransformer("BAAI/bge-small-en-v1.5", device=device)
self.ce = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2", device=device)
# KEYWORD arm. Preferred: prebuilt disk-backed FTS5 chunk index (escr_fts.sqlite, built
# offline by build_fts_index.py) β ms queries, ~0 resident RAM, no boot cost. Fallback:
# the in-RAM doc-level BM25 build (THEMIS_KEYWORD=1). THEMIS_FTS=0 opts out of FTS5.
self.kw = None; self.fts = None
_fts_path = os.path.join(data_dir, "escr_fts.sqlite")
if os.environ.get("THEMIS_FTS", "1") == "1" and os.path.exists(_fts_path):
import sqlite3, threading
self.fts = sqlite3.connect(f"file:{_fts_path}?mode=ro", uri=True, check_same_thread=False)
self.fts_lock = threading.Lock() # lanes hit keyword_search from parallel threads
print(f"[tools] FTS5 keyword index: {os.path.getsize(_fts_path)/1e9:.2f} GB (disk-backed)", flush=True)
if self.fts is None and os.environ.get("THEMIS_KEYWORD", "1") == "1":
import math, time as _t
t0 = _t.time(); print("[tools] building keyword index ...", flush=True)
self.kw_docs = list(self.doc_chunks.keys())
self.kw_postings = defaultdict(list); df = Counter()
self.kw_dl = np.zeros(len(self.kw_docs), dtype=np.float32)
for i, d in enumerate(self.kw_docs):
toks = []
for ci in self.doc_chunks[d]: toks += re.findall(r"[a-z0-9]+", self.texts[ci].lower())
tf = Counter(toks); self.kw_dl[i] = len(toks)
for t, c in tf.items(): self.kw_postings[t].append((i, c)); df[t] += 1
N = len(self.kw_docs); self.kw_avgdl = float(self.kw_dl.mean()) or 1.0
self.kw_idf = {t: math.log(1 + (N - n + 0.5) / (n + 0.5)) for t, n in df.items()}
self.kw = True
print(f"[tools] keyword index: {len(self.kw_postings)} terms, {N} docs, {_t.time()-t0:.0f}s", flush=True)
print(
f"[tools] ready β {len(self.eligible_doc_ids)} source-grounded judgments "
f"({len(self.meta) - len(self.eligible_doc_ids)} metadata-only excluded), "
f"{len(self.statute_idx)} statute sections",
flush=True,
)
# ---- helpers ----
def _enc(self, q):
return self.st.encode(BGE_Q + q, normalize_embeddings=True, convert_to_numpy=True).astype(np.float32)
def is_retrieval_eligible(self, doc_id):
"""True only for judgments whose source text is in the active corpus."""
return str(doc_id) in self.eligible_doc_ids
def coverage(self):
return {
"accepted_judgments": len(self.eligible_doc_ids),
"metadata_only_excluded": len(self.meta) - len(self.eligible_doc_ids),
"scope": "Supreme Court of India judgments stored in this release",
}
def _card(self, d, rr=0.0):
m = self.meta.get(d, {}); gl = self.goodlaw.get(d, {})
cis = self.doc_chunks.get(d, [])
snip = _clean((m.get("held") or m.get("issue")
or (self.syn_held.get(d) or {}).get("held")
or (self.texts[cis[0]] if cis else "")))[:240]
return {"doc_id": d, "judgment_id": str(d), "case_name": m.get("case_name"),
"year": self.decision_year.get(d) or m.get("year") or m.get("date"),
"neutral_citation": m.get("neutral_citation"), "bench_strength": m.get("bench_strength"),
"cited_by": self.cite_indeg.get(d, 0), "good_law": gl.get("good_law_status", "unknown"),
"rr": round(float(rr), 2), "snippet": snip}
def _dense_pool(self, qv, n=120):
sim = self.M @ qv
requested = min(max(n * 2, n + 1), len(sim) - 1)
top = np.argpartition(-sim, requested)[:requested]
return [
int(i) for i in top[np.argsort(-sim[top])]
if self.is_retrieval_eligible(self.chunk_doc[int(i)])
][:n]
def _rerank_docs(self, q, cand_idx, topk):
cand_idx = [
ci for ci in cand_idx
if self.is_retrieval_eligible(self.chunk_doc[ci])
and _clean(self.texts[ci])
]
if not cand_idx:
return []
rr = self.ce.predict([(q, self.texts[ci]) for ci in cand_idx])
best = {}
for ci, s in zip(cand_idx, rr):
d = self.chunk_doc[ci]
if not self.is_retrieval_eligible(d): continue
if d not in best or s > best[d]: best[d] = float(s)
ranked = sorted(best.items(), key=lambda x: -x[1])[:topk]
return [self._card(d, s) for d, s in ranked]
# ---- TOOLS ----
def vector_search(self, q, k=8):
"""Semantic retrieval β doctrine described in the user's words. dense pool -> cross-encoder."""
return self._rerank_docs(q, self._dense_pool(self._enc(q), 120), k)
def authority_search(self, q, k=8, alpha=0.3):
"""Retrieve, then rank by AUTHORITY (cross-encoder + alpha*log1p(cite_indeg)) β 'the leading case on X'."""
cand = self._dense_pool(self._enc(q), 120)
rr = self.ce.predict([(q, self.texts[ci]) for ci in cand]); best = {}
for ci, s in zip(cand, rr):
d = self.chunk_doc[ci]
if not self.is_retrieval_eligible(d): continue
if d not in best or s > best[d]: best[d] = float(s)
sig = lambda x: 1/(1+np.exp(-x))
scored = sorted(best.items(), key=lambda x: -(sig(x[1]) + alpha*np.log1p(self.cite_indeg.get(x[0], 0))))[:k]
return [self._card(d, s) for d, s in scored]
def keyword_search(self, q, k=12, k1=1.5, b=0.75):
"""BM25 keyword retrieval β exact terms / names / section nums that dense misses.
FTS5 path: chunk-level match, best-chunk-per-doc aggregation (a doc with one strong
exact-term chunk ranks high), OR semantics to mirror the legacy scorer."""
if self.fts is not None:
_stop = {"of","the","and","or","in","to","a","an","is","for","on","by","at",
"with","under","was","were","be","has","had","it","that","this"}
toks = [t for t in re.findall(r"[a-z0-9]+", q.lower()) if t not in _stop]
if not toks: return []
match = " OR ".join(f'"{t}"' for t in toks[:24]) # quoted: immune to FTS syntax chars
with self.fts_lock:
rows = self.fts.execute(
"SELECT rowid, bm25(fts) FROM fts WHERE fts MATCH ? ORDER BY bm25(fts) LIMIT ?",
(match, max(k * 12, 240))).fetchall()
best = {}
for ci, s in rows: # bm25(): smaller = better
d = self.chunk_doc[ci]
if not self.is_retrieval_eligible(d): continue
if d not in best or s < best[d]: best[d] = s
top = sorted(best.items(), key=lambda x: x[1])[:k]
return [self._card(d) for d, _ in top]
if not self.kw: return []
scores = defaultdict(float)
for t in set(re.findall(r"[a-z0-9]+", q.lower())):
idf = self.kw_idf.get(t)
if not idf: continue
for i, tf in self.kw_postings[t]:
scores[i] += idf * (tf * (k1 + 1)) / (tf + k1 * (1 - b + b * self.kw_dl[i] / self.kw_avgdl))
top = [
item for item in sorted(scores.items(), key=lambda x: -x[1])
if self.is_retrieval_eligible(self.kw_docs[item[0]])
][:k]
return [self._card(self.kw_docs[i]) for i, _ in top]
def dense_docs(self, q, k=60):
"""Doc-level dense ranking (first chunk-hit per doc)."""
qv = self._enc(q); sim = self.M @ qv
top = np.argpartition(-sim, 2500)[:2500]; top = top[np.argsort(-sim[top])]
seen = []; s = set()
for ci in top:
d = self.chunk_doc[ci]
if not self.is_retrieval_eligible(d): continue
if d not in s: s.add(d); seen.append(d)
if len(seen) >= k: break
return seen
def held_search(self, q, k=12):
"""Rank judgments by HELD-headnote similarity (doc-level, clean reporter language)."""
if self.held_V is None: return []
qv = self._enc(q); sim = self.held_V @ qv
top = np.argpartition(-sim, min(k, len(sim) - 1))[:k]
return [
self.held_docs[int(i)] for i in top[np.argsort(-sim[top])]
if self.is_retrieval_eligible(self.held_docs[int(i)])
][:k]
def citectx_search(self, q, k=12):
"""Rank judgments by how LATER courts describe them (citation-context vectors).
Multiple contexts per doc -> dedupe keeping best rank."""
if self.citectx_V is None: return []
qv = self._enc(q); sim = np.asarray(self.citectx_V @ qv)
n = min(k * 6, len(sim) - 1)
top = np.argpartition(-sim, n)[:n]
out, seen = [], set()
for i in top[np.argsort(-sim[top])]:
d = self.citectx_docs[int(i)]
if not self.is_retrieval_eligible(d): continue
if d not in seen:
seen.add(d); out.append(d)
if len(out) >= k: break
return out
def hybrid_search(self, q, k=8, pool=60):
"""The strong base primitive (panel + recall ablation: RRF@100=0.96): dense + BM25 -> RRF -> rerank."""
dd = self.dense_docs(q, pool)
kd = [c["doc_id"] for c in self.keyword_search(q, pool)]
sc = {}
for r in (dd, kd):
for rank, d in enumerate(r): sc[d] = sc.get(d, 0.0) + 1.0 / (60 + rank + 1)
fused = [d for d, _ in sorted(sc.items(), key=lambda x: -x[1])][:max(40, k * 4)]
rr = self.score_docs(q, fused)
for d in fused:
if d not in rr: rr[d] = -9.0
ranked = sorted(fused, key=lambda d: -rr[d])
return [self._card(d, rr.get(d, 0.0)) for d in ranked[:k]]
def statute_search(self, q, k=3):
"""Find the statute SECTION(S) most relevant to the query (BNS/IPC/CrPC/IEA/...)."""
qv = self._enc(q); sim = self.statute_V @ qv
out = []
for j in np.argsort(-sim)[:k]:
s = self.statute_idx[int(j)]
out.append({"act": s.get("act_short"), "section": s.get("section_number"),
"title": s.get("title"), "i": int(j)})
return out
def cases_on_section(self, act_section_text, k=8):
"""Cases discussing a statute section: embed the section text, retrieve nearest judgments."""
qv = self.st.encode(act_section_text, normalize_embeddings=True, convert_to_numpy=True).astype(np.float32)
return self._rerank_docs(act_section_text[:300], self._dense_pool(qv, 120), k)
def statute_crosswalk(self, code, section):
"""Map a section across the new/old codes (BNS<->IPC, BNSS<->CrPC, BSA<->IEA)."""
return self.crosswalk.lookup(code, section)
def statute_provision(self, code, section):
return self.statute_library.lookup(code, section)
def encode_documents(self, texts):
values = [_clean(value) for value in texts if _clean(value)]
if not values:
return np.empty((0, int(self.M.shape[1])), dtype=np.float32)
return np.asarray(
self.st.encode(values, normalize_embeddings=True, convert_to_numpy=True),
dtype=np.float32,
)
def find_similar_cases(self, doc_id, k=8):
"""'More like this' β nearest judgments to doc_id by embedding centroid."""
cis = self.doc_chunks.get(doc_id, [])
if not cis: return []
centroid = self.M[cis].mean(0); centroid /= (np.linalg.norm(centroid) + 1e-9)
pool = self._dense_pool(centroid.astype(np.float32), 60)
seen = set([doc_id]); out = []
for ci in pool:
d = self.chunk_doc[ci]
if self.is_retrieval_eligible(d) and d not in seen:
seen.add(d); out.append(self._card(d))
if len(out) >= k: break
return out
def cited_authorities(self, doc_id, k=12):
"""Note-UP: the cases doc_id relies on (its authority chain)."""
return [
self._card(d)
for d in list(dict.fromkeys(self.out_edges.get(doc_id, [])))
if self.is_retrieval_eligible(d)
][:k]
def progeny(self, doc_id, k=12):
"""Note-DOWN: the cases that cite doc_id (its progeny + treatment)."""
out = []
for d in list(dict.fromkeys(self.in_edges.get(doc_id, [])))[:k]:
if not self.is_retrieval_eligible(d): continue
c = self._card(d); c["treatment"] = self.edge_meta.get((d, doc_id), {}).get("treatment")
out.append(c)
return out
def co_cited_cases(self, doc_id, k=8):
"""Cases similar by SHARED AUTHORITIES (bibliographic coupling) β cases that cite what doc_id cites."""
mine = set(self.out_edges.get(doc_id, []))
if not mine: return []
score = Counter()
for t in mine:
for citer in self.in_edges.get(t, []):
if citer != doc_id and self.is_retrieval_eligible(citer): score[citer] += 1
return [self._card(d) for d, _ in score.most_common(k)]
def good_law_check(self, doc_id):
"""Citator: current status + treatment breakdown + the overruling case if any."""
gl = self.goodlaw.get(doc_id, {})
status = gl.get("good_law_status", "unknown")
overruled_by = None
if status in BAD_STATUS:
for s, t in [(s, t) for (s, t) in self.edge_meta if t == doc_id]:
if self.edge_meta[(s, t)].get("treatment") in ("overruled", "overrules"):
if self.is_retrieval_eligible(s):
overruled_by = self._card(s); break
return {"doc_id": doc_id, "good_law": status, "treatment_breakdown": gl.get("treatment_breakdown", {}),
"overruled_by": overruled_by}
def metadata_filter(self, cards, min_bench=None, year_from=None, year_to=None):
"""Filter a candidate list by bench strength (Constitution Bench = 5+), date range."""
out = []
_BN = {"single": 1, "division": 2, "full": 3, "constitution": 5, "larger": 7}
for c in cards:
d = c["doc_id"]; m = self.meta.get(d, {})
bs = self.bench_n.get(d) or _BN.get(str(m.get("bench_strength") or "").lower(), 0)
yr = self.decision_year.get(d) or 0
if not yr:
try: yr = int(str(m.get("year") or 0)[:4])
except Exception: yr = 0
if min_bench and bs < min_bench: continue
if year_from and yr and yr < year_from: continue
if year_to and yr and yr > year_to: continue
out.append(c)
return out
def read_case(self, doc_id):
"""Read a case's headnote/held/issue (for the agent to verify relevance + for grounding)."""
if not self.is_retrieval_eligible(doc_id):
return {}
m = self.meta.get(doc_id, {})
held = _clean(m.get("held")); issue = _clean(m.get("issue"))
if not held and doc_id in self.syn_held: # LLM-backfilled reporter-style headnote
s = self.syn_held[doc_id]
held = _clean(s.get("held")); issue = issue or _clean(s.get("issue"))
if not held and not issue: # older cases lack extracted headnotes -> fall back to first chunks
cis = self.doc_chunks.get(doc_id, [])
held = _clean(" ".join(self.texts[i] for i in cis[:2]))
return {"doc_id": doc_id, "case_name": m.get("case_name"), "neutral_citation": m.get("neutral_citation"),
"bench_strength": m.get("bench_strength"), "good_law": self.goodlaw.get(doc_id, {}).get("good_law_status", "unknown"),
"issue": issue[:1200], "held": held[:1800]}
def score_docs(self, q, doc_ids, per_doc=3):
"""Uniformly cross-encoder-score a heterogeneous pool of docs vs q (max over each doc's first
chunks). One batched CE pass. Returns {doc_id: rr}. Lets graph/authority/statute additions be
ranked on the same scale as dense hits."""
pairs = []; owner = []
for d in doc_ids:
if not self.is_retrieval_eligible(d): continue
for ci in self.doc_chunks.get(d, [])[:per_doc]:
pairs.append((q, self.texts[ci])); owner.append(d)
if not pairs: return {d: 0.0 for d in doc_ids}
sc = self.ce.predict(pairs, batch_size=256)
best = {d: -9e9 for d in doc_ids}
for d, s in zip(owner, sc):
if s > best[d]: best[d] = float(s)
return {d: (best[d] if best[d] > -9e9 else 0.0) for d in doc_ids}
def front_text(self, doc_id, n=1800):
"""The judgment's FRONT MATTER (reporter headnote lives here in 70-100% of judgments across
all decades β more reliable than meta.held, which craters to ~2% in the 1970s-80s)."""
held = _clean(self.meta.get(doc_id, {}).get("held") or "")
if len(held) > 200: return held[:n]
syn = _clean((self.syn_held.get(doc_id) or {}).get("held") or "")
if len(syn) > 200: return syn[:n] # backfilled crater doc: skim the ratio, not the cover page
cis = self.doc_chunks.get(doc_id, [])
return _clean(" ".join(self.texts[i] for i in cis[:3]))[:n]
def full_text_for_read(self, q, doc_id, cap_chars=90000):
"""Layer-2 reading surface: the FULL judgment up to ~cap (β22k tokens). Above-cap monsters get a
tiered pack: HELD + opening + a window around the query's best-matching chunk + the ending β
the controlling passage in multi-issue judgments sits mid-text where head/tail packs go blind."""
cis = self.doc_chunks.get(doc_id, [])
if not cis: return ""
parts = [self.texts[i] for i in cis]
full = "\n".join(parts)
if len(full) <= cap_chars: return full
held = _clean((self.meta.get(doc_id, {}).get("held") or ""))[:6000]
probe = cis[:40] # find the query-relevant window (one small CE pass)
sc = self.ce.predict([(q, self.texts[i]) for i in probe])
bi = int(np.argmax(sc))
win = "\n".join(self.texts[i] for i in cis[max(0, bi - 2):bi + 3])
head = "\n".join(parts[:8]); tail = "\n".join(parts[-6:])
pack = (("HELD: " + held + "\n\n") if held else "") + head + "\n[...]\n" + win + "\n[...]\n" + tail
return pack[:cap_chars]
def best_chunk_text(self, q, doc_id, limit=1600):
"""The doc's single chunk most relevant to q (for the grounding gate to quote from)."""
cis = self.doc_chunks.get(doc_id, [])[:6]
if not cis: return ""
sc = self.ce.predict([(q, self.texts[ci]) for ci in cis])
return _clean(self.texts[cis[int(np.argmax(sc))]])[:limit]
def judgment_view(self, doc_id):
"""Full case view for the pilot UI (metadata + issue/held + good-law + cited-by). issue is in
~3% of metadata, held in ~56% β so fall back to the judgment's opening text when missing."""
if not self.is_retrieval_eligible(doc_id):
return {}
m = self.meta.get(doc_id, {}); gl = self.goodlaw.get(doc_id, {})
cis = self.doc_chunks.get(doc_id, [])
body = _clean(" ".join(self.texts[i] for i in cis[:14]))
extracted = self.case_summaries.get(doc_id, {})
issue = _clean(m.get("issue") or extracted.get("issue"))
held = _clean(m.get("held") or extracted.get("held")); synthetic = False
if not held and doc_id in self.syn_held: # LLM-backfilled headnote (disclosed to the UI)
s = self.syn_held[doc_id]
held = _clean(s.get("held")); issue = issue or _clean(s.get("issue")); synthetic = bool(held)
if not held: held = body[:6000] # no extracted headnote -> show the opening (the headnote lives there)
return {"doc_id": doc_id, "synthetic_headnote": synthetic,
"summary": case_summary_record(m, self.syn_held.get(doc_id), extracted),
"case_name": m.get("case_name"), "neutral_citation": m.get("neutral_citation"),
"equivalent_citations": m.get("equivalent_citations"), "court": m.get("court"), "date": m.get("date"),
"bench_strength": m.get("bench_strength"), "disposition": m.get("disposition"),
"good_law_status": gl.get("good_law_status", "unknown"), "treatment_breakdown": gl.get("treatment_breakdown", {}),
"cited_by": self.cite_indeg.get(doc_id, 0), "issue": issue[:4000],
"held": held[:6000], "text": body[:60000]}
def case_chat_passages(self, q, doc_id, k=4, limit=1800):
"""Return only source passages from one eligible opened judgment.
The legacy bundle has chunk identity rather than schema-v5 paragraph
identity, so its stable fallback IDs are disclosed as chunk anchors.
The v5 adapter supplies real paragraph IDs through the same contract.
"""
if not self.is_retrieval_eligible(doc_id):
return []
cis = self.doc_chunks.get(doc_id, [])
if not cis:
return []
scores = self.ce.predict([(q, self.texts[ci]) for ci in cis[:40]])
ranked = sorted(zip(cis[:40], scores), key=lambda item: -float(item[1]))[:k]
return [
{
"paragraph_id": f"{doc_id}:chunk:{ci}",
"label": f"Indexed passage {rank + 1}",
"text": _clean(self.texts[ci])[:limit],
"source_kind": "legacy_chunk",
}
for rank, (ci, _) in enumerate(ranked, 1)
if _clean(self.texts[ci])
]
def identity_hits(self, q):
"""Known-item route: a citation or a 'X v Y' case-name query resolves to the EXACT case(s)
(cite_indeg salience tiebreak), not semantic search. Restores serve.py's identity routing."""
ql = q.strip()
m = re.search(r"\[\d{4}\]\s*\d+\s*S\.?C\.?R\.?\s*\d+|\(\d{4}\)\s*\d+\s*SCC\s*\d+|\d{4}\s+INSC\s+\d+|AIR\s+\d{4}\s+SC\s+\d+", ql, re.I)
if m:
rid = self.cite_resolver.get(re.sub(r"\s+", " ", m.group(0).replace(".", "")).strip().upper()) or self.nc2doc.get(m.group(0))
if rid and self.is_retrieval_eligible(rid): return [rid], "citation"
# famous-name alias ("kesavananda", "the shah bano judgment") β exact or contained phrase
if self.aliases and len(ql) <= 60:
qa = re.sub(r"[^a-z0-9 ]", " ", ql.lower())
qa = re.sub(r"\b(the|case|judgment|judgement|in|re|of)\b", " ", qa)
qa = re.sub(r"\s+", " ", qa).strip()
if qa in self.aliases and self.is_retrieval_eligible(self.aliases[qa]):
return [self.aliases[qa]], "case name"
# substring form ("the kesavananda bharati judgment") β but ONLY when the query is
# essentially just the name: doctrinal residue ("bachan singh sentencing principles")
# must fall through to full retrieval, not the single-doc shortcut
hits = [(a, d) for a, d in self.aliases.items()
if len(a) >= 8 and a in qa and len(qa) - len(a) <= 10]
if hits:
best = max(hits, key=lambda ad: self.cite_indeg.get(ad[1], 0))
if self.is_retrieval_eligible(best[1]):
return [best[1]], "case name"
if re.search(r"\bv[s.]?\b|\bversus\b", ql, re.I) and len(ql) <= 90:
hits = [c["doc_id"] for c in self.name_lookup(ql, 6)]
if hits: return hits, "case name"
return [], None
def name_lookup(self, name, k=4):
"""Resolve a case NAME to corpus doc(s) β the recall tool for LLM-named authorities."""
raw = [t for t in _ntok(name) if t not in _NAME_STOP and len(t) > 1]
if not raw: return []
# compound-name variants (Indian names split/join freely: Ibrahimuddin <-> Ibrahim Uddin)
extra = []
for t in raw:
if t not in self.name_vocab and len(t) >= 7: # try splitting an unknown long token
for cut in range(3, len(t) - 2):
a, b = t[:cut], t[cut:]
if a in self.name_vocab and b in self.name_vocab: extra += [a, b]; break
for a, b in zip(raw, raw[1:]): # try joining adjacent tokens
if (a + b) in self.name_vocab: extra.append(a + b)
raw += extra
qtok = set()
for t in raw:
if t in self.name_vocab or len(t) <= 3: qtok.add(t)
else: qtok.update(difflib.get_close_matches(t, self.name_vocab, n=3, cutoff=0.82) or [t])
cand = set() # only docs sharing a query token (inverted index)
for t in qtok: cand |= self.name_postings.get(t, set())
qdist = {t for t in qtok if len(t) >= 5} # distinctive party-name tokens (must match one)
scored = []
for d in cand:
if not self.is_retrieval_eligible(d): continue
ntok = set(_ntok(self.meta.get(d, {}).get("case_name") or ""))
ov = qtok & ntok
if qdist and not (qdist & ntok): continue # reject namesakes that miss the party name
if len(ov) >= 2 or (len(ov) == 1 and any(len(t) >= 5 for t in ov)):
# rank: most query tokens matched, then the sibling-cluster CANONICAL (the main
# judgment, not its referral order), then concision, then authority
scored.append((len(ov), 1 if d in self.canonical else 0,
-(len(ntok) - len(ov)), self.cite_indeg.get(d, 0), d))
scored.sort(reverse=True)
return [self._card(t[-1]) for t in scored[:k]]
|