File size: 6,146 Bytes
4e1037f | 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 | """DAIL-style few-shot selection helpers (phase A2).
Default product path embeds the *raw* question against the ``fewshot_qsql``
collection (dense top-k). DAIL-SQL (arXiv:2308.15363) and the MCS-SQL
ablation instead embed a *schema-masked* question so retrieval prefers
intent/structure over shared table/column names.
This module implements the cheap **2a** variant only: mask schema tokens
in the query text before embedding. Skeleton re-ranking (2b) is deferred
until 2a shows ≥ +1.0 pp on n=200.
Default OFF — wired only when ``PipelineConfig.fewshot_selection == "dail"``.
"""
from __future__ import annotations
import re
from collections.abc import Iterable
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from nl_sql.schema_index.indexer import SchemaIndex
MASK_TOKEN = "<mask>"
_MIN_TOKEN_LEN = 2
# Column line in chunker-rendered table cards:
# " - District Name: TEXT [NULL] | nulls=0 ..."
_COLUMN_LINE = re.compile(r"^\s+-\s+([^:]+):\s+\S+", re.MULTILINE)
# Ultra-generic identifiers that are also ordinary English; masking them
# shreds the question into noise. Table names are never filtered this way.
_GENERIC_IDENTIFIERS = frozenset(
{
"id",
"name",
"type",
"date",
"time",
"year",
"month",
"day",
"code",
"key",
"value",
"text",
"data",
"info",
"number",
"count",
"total",
"status",
"level",
"order",
"group",
"user",
"index",
"rank",
"score",
"rate",
"flag",
"mode",
"kind",
"class",
"size",
"start",
"end",
"first",
"last",
"min",
"max",
"avg",
"sum",
"desc",
"description",
"comment",
"note",
"notes",
"title",
"label",
"amount",
"price",
"cost",
"qty",
"quantity",
"age",
"sex",
"gender",
"city",
"state",
"country",
"address",
"phone",
"email",
"url",
"path",
"file",
}
)
def parse_column_names_from_chunk(text: str) -> set[str]:
"""Extract column identifiers from a chunker-rendered table card."""
names: set[str] = set()
for match in _COLUMN_LINE.finditer(text or ""):
raw = match.group(1).strip()
if raw:
names.add(raw)
return names
def collect_schema_tokens(index: SchemaIndex, db_id: str) -> frozenset[str]:
"""Table + column names for ``db_id`` from the indexed schema cards.
Used only as a query-time mask source — does not rebuild embeddings.
"""
if not db_id:
return frozenset()
records = index.schema_collection.get(
where={"db_id": db_id},
include=["documents", "metadatas"],
)
tokens: set[str] = set()
for meta in records.get("metadatas") or []:
if not meta:
continue
table = str(meta.get("table_name") or "").strip()
if table:
tokens.add(table)
for doc in records.get("documents") or []:
tokens.update(parse_column_names_from_chunk(str(doc or "")))
return frozenset(t.strip() for t in tokens if _keep_token(t))
def _keep_token(token: str) -> bool:
t = token.strip()
if len(t) < _MIN_TOKEN_LEN:
return False
# Pure digits / single symbols are never useful schema anchors.
if t.isdigit():
return False
folded = t.casefold()
if folded in _GENERIC_IDENTIFIERS:
return False
# Multi-word / underscored names always keep (e.g. "District Name").
if " " in t or "_" in t:
return True
return True
def _expand_variants(token: str) -> list[str]:
"""Surface space/underscore variants so NL phrasing still masks."""
base = token.strip()
if not base:
return []
variants = {base, base.replace("_", " "), base.replace(" ", "_")}
# Prefer longest first so "District Name" wins over "District".
return sorted(variants, key=len, reverse=True)
def mask_schema_tokens(
question: str,
tokens: Iterable[str],
*,
mask: str = MASK_TOKEN,
) -> str:
"""Replace whole-token schema identifiers in ``question`` with ``mask``.
Matching is case-insensitive. Longer tokens are applied first so multi-word
column names are not partially eaten by a shorter component. Tokens that
never appear leave the question unchanged (cheap no-op path).
"""
if not question or not tokens:
return question
# Dedup expanded variants, longest first.
seen: set[str] = set()
ordered: list[str] = []
for tok in tokens:
for variant in _expand_variants(str(tok)):
key = variant.casefold()
if key in seen or not _keep_token(variant):
continue
# Re-check generic on the variant itself.
if (
variant.casefold() in _GENERIC_IDENTIFIERS
and " " not in variant
and "_" not in variant
):
continue
seen.add(key)
ordered.append(variant)
ordered.sort(key=len, reverse=True)
result = question
for tok in ordered:
escaped = re.escape(tok)
# Word-ish boundaries: not alphanumeric on either side. Works for
# "Album", "District Name", and "frpm" without eating "frpm_table".
pattern = re.compile(rf"(?<![A-Za-z0-9]){escaped}(?![A-Za-z0-9])", re.IGNORECASE)
result = pattern.sub(mask, result)
# Collapse accidental double masks from overlapping replacements.
return re.sub(rf"(?:{re.escape(mask)}\s*){{2,}}", f"{mask} ", result)
def fewshot_query_text(
question: str,
*,
selection: str,
schema_tokens: Iterable[str],
) -> str:
"""Return the text that should be embedded for few-shot retrieval."""
mode = (selection or "dense").strip().lower()
if mode == "dail":
return mask_schema_tokens(question, schema_tokens)
return question
|