feat(style-acq T1-T3): קורפוס-דוגמאות של דפנה לכותב (style_exemplars)

ממלא את ערוץ-הדוגמאות (B) של מערכת רכישת-הסגנון: הכותב מאחזר פסקאות-בלוק
אמיתיות של דפנה בזמן כתיבה, ממוקדות section+outcome+practice_area.

T1 — תשתית + backfill:
- SCHEMA_V27: טבלת style_exemplars (purpose-built — בלי תיקים מזויפים בשרשרת
  decision_paragraphs). decision_number/source/section/outcome/practice_area+embedding.
- db: insert/delete/search_style_exemplars + count_style_exemplars.
- scripts/backfill_style_exemplars.py: מפצל קורפוס דפנה (style_corpus +
  internal_committee) לסעיפים→פסקאות, embed, שמירה. אידמפוטנטי, dry-run/apply.

T2 — אחזור ממוקד:
- search_style_exemplars(section, outcome, practice_area) — section=hard filter,
  outcome/practice_area=soft. block_writer._build_precedents_context ממפה
  block→section ומאחזר (ראשי), לצד הנתיב הישן (משלים).

T3 — contrastive/adapt:
- הדוגמאות מתויגות "מבנה/קול בלבד — התאם, אל תעתיק תוכן"; פסקה מלאה (1100 תווים).

INV-LRN5 (טוהר — סגנון בלבד). G11. הרצת backfill --apply בנפרד.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-06-06 18:10:01 +00:00
parent a3451775fa
commit 2e20e27e17
4 changed files with 261 additions and 3 deletions

View File

@@ -725,19 +725,43 @@ async def _build_precedents_context(
style_parts: list[str] = []
caselaw_parts: list[str] = []
case_law_ids: list[str] = []
# block → golden-ratio section, for targeted exemplar retrieval (T2)
_BLOCK_SECTION = {
"block-vav": "background", "block-zayin": "claims",
"block-yod": "discussion", "block-yod-alef": "summary",
}
try:
case = await db.get_case(case_id)
case_number = case.get("case_number", "") if case else ""
subject = case.get("subject", "") if case else ""
practice_area = case.get("practice_area", "") if case else ""
decision = await db.get_decision_by_case(case_id)
outcome = (decision or {}).get("outcome", "")
query = f"דיון משפטי בנושא {subject}" if subject else "דיון משפטי ועדת ערר"
query_emb = await embeddings.embed_query(query)
section = _BLOCK_SECTION.get(block_id)
# Stream 1: paragraph_embeddings — Dafna's own prose (STYLE exemplars, not content)
# Stream 1a (PRIMARY): Dafna's own block-level prose from her corpus
# (style_exemplars) — matched by section + outcome + practice_area (T2/T3).
if section:
exemplars = await db.search_style_exemplars(
query_embedding=query_emb, section=section,
outcome=outcome or None, practice_area=practice_area or None, limit=6,
)
exemplars = [e for e in exemplars if e.get("decision_number", "") != case_number]
for e in exemplars[:4]:
style_parts.append(
f"[דוגמת-סגנון (מבנה/קול בלבד — התאם, אל תעתיק תוכן) — "
f"{e.get('decision_number', '?')}, {section}, "
f"outcome={e.get('outcome') or ''}]\n{e['paragraph_text'][:1100]}"
)
# Stream 1b: paragraphs from pipeline cases (legacy path; may be empty)
para_results = await db.search_similar_paragraphs(
query_embedding=query_emb, limit=10, block_type="block-yod",
)
para_results = [r for r in para_results if r.get("case_number", "") != case_number]
for r in para_results[:4]:
for r in para_results[:2]:
style_parts.append(
f"[דוגמת-סגנון — החלטת {r.get('case_number', '?')} "
f"{r.get('case_title', '')}, בלוק {r.get('block_type', '')}]\n{r['content'][:500]}"

View File

@@ -1204,6 +1204,28 @@ CREATE INDEX IF NOT EXISTS idx_draft_final_pairs_case ON draft_final_pairs(case_
CREATE INDEX IF NOT EXISTS idx_draft_final_pairs_status ON draft_final_pairs(status);
"""
SCHEMA_V27_SQL = """
-- style_exemplars (T1-T3): block-level paragraphs from Dafna's OWN decisions
-- (style_corpus + internal_committee finals), embedded for retrieval as
-- style exemplars at write-time. Purpose-built so we DON'T fabricate synthetic
-- cases just to reuse decision_paragraphs. INV-LRN5: style material only — the
-- writer is told to adapt structure/voice, copy only boilerplate, never substance.
CREATE TABLE IF NOT EXISTS style_exemplars (
id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
decision_number TEXT DEFAULT '',
source TEXT DEFAULT '', -- style_corpus | internal_committee
practice_area TEXT DEFAULT '',
outcome TEXT DEFAULT '', -- rejection | partial_acceptance | full_acceptance | ''
section TEXT DEFAULT 'other', -- background | claims | discussion | summary | other
paragraph_text TEXT NOT NULL,
word_count INTEGER DEFAULT 0,
embedding vector(1024),
created_at TIMESTAMPTZ DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_style_exemplars_section ON style_exemplars(section);
CREATE INDEX IF NOT EXISTS idx_style_exemplars_decision ON style_exemplars(decision_number, source);
"""
async def _run_schema_migrations(pool: asyncpg.Pool) -> None:
async with pool.acquire() as conn:
@@ -1234,7 +1256,8 @@ async def _run_schema_migrations(pool: asyncpg.Pool) -> None:
await conn.execute(SCHEMA_V24_SQL)
await conn.execute(SCHEMA_V25_SQL)
await conn.execute(SCHEMA_V26_SQL)
logger.info("Database schema initialized (v1-v26)")
await conn.execute(SCHEMA_V27_SQL)
logger.info("Database schema initialized (v1-v27)")
async def init_schema() -> None:
@@ -2329,6 +2352,85 @@ async def list_draft_final_pairs(status: str | None = None, limit: int = 200) ->
return [dict(r) for r in rows]
async def insert_style_exemplar(
decision_number: str, source: str, practice_area: str, outcome: str,
section: str, paragraph_text: str, word_count: int, embedding: list[float],
) -> None:
"""Insert one block-level style exemplar (T1 backfill)."""
pool = await get_pool()
async with pool.acquire() as conn:
await conn.execute(
"""INSERT INTO style_exemplars
(decision_number, source, practice_area, outcome, section,
paragraph_text, word_count, embedding)
VALUES ($1, $2, $3, $4, $5, $6, $7, $8)""",
decision_number, source, practice_area, outcome, section,
paragraph_text, word_count, str(embedding),
)
async def delete_style_exemplars(decision_number: str, source: str) -> int:
"""Idempotent backfill: clear a decision's exemplars before re-inserting."""
pool = await get_pool()
async with pool.acquire() as conn:
res = await conn.execute(
"DELETE FROM style_exemplars WHERE decision_number = $1 AND source = $2",
decision_number, source,
)
try:
return int(res.split()[-1])
except (ValueError, IndexError):
return 0
async def search_style_exemplars(
query_embedding: list[float],
section: str | None = None,
outcome: str | None = None,
practice_area: str | None = None,
limit: int = 6,
) -> list[dict]:
"""Retrieve Dafna's own block-level paragraphs as STYLE exemplars (T2).
Filters by section (block) + optionally outcome/practice_area for the closest
match to the block being written. Soft filters: outcome/practice_area narrow but
never zero-out — section is the hard filter."""
pool = await get_pool()
conditions, params, idx = [], [query_embedding, limit], 3
if section:
conditions.append(f"section = ${idx}"); params.append(section); idx += 1
if outcome:
conditions.append(f"(outcome = ${idx} OR outcome = '')"); params.append(outcome); idx += 1
if practice_area:
conditions.append(f"(practice_area = ${idx} OR practice_area = '')"); params.append(practice_area); idx += 1
where = f"WHERE {' AND '.join(conditions)}" if conditions else ""
sql = f"""
SELECT decision_number, source, section, outcome, practice_area,
paragraph_text, word_count,
1 - (embedding <=> $1) AS score
FROM style_exemplars
{where}
ORDER BY embedding <=> $1
LIMIT $2
"""
async with pool.acquire() as conn:
rows = await conn.fetch(sql, *params)
return [dict(r) for r in rows]
async def count_style_exemplars() -> dict:
"""Coverage check for the backfill."""
pool = await get_pool()
async with pool.acquire() as conn:
total = await conn.fetchval("SELECT count(*) FROM style_exemplars")
by_section = await conn.fetch(
"SELECT section, count(*) AS n FROM style_exemplars GROUP BY section ORDER BY n DESC"
)
decisions = await conn.fetchval(
"SELECT count(DISTINCT decision_number) FROM style_exemplars"
)
return {"total": total, "decisions": decisions, "by_section": [dict(r) for r in by_section]}
async def upsert_style_pattern(
pattern_type: str,
pattern_text: str,