File size: 12,875 Bytes
458c8e2 |
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 |
from __future__ import annotations
import json
from typing import Any, Dict, List, Optional
from langchain.tools import tool
from langchain_community.tools.tavily_search import TavilySearchResults
from pipeline.deduplication import find_near_duplicates
from pipeline.featurizer import custom_featurizer
from pipeline.issues import find_issues
from pipeline.utils_cool import (
df_from_payload,
df_to_payload,
get_signature_dict,
guess_task_and_label,
)
from .runtime_ctx import (
get_df_payload, # now supports version spec (None|'current'|'prev'|'base'|'@-1'|int)
get_df_summary,
get_sota_bundled,
set_df_payload, # commit new dataset version (or replace)
set_df_summary,
set_sota_bundled,
)
from .runtime_ctx import (
list_versions as _list_versions_state,
)
from .runtime_ctx import (
reset_current_to as _reset_current_to,
)
# Registry of runnable steps (names used by the agent/UI)
STEP_FUNCS = {
"dedup": find_near_duplicates,
"featurize": custom_featurizer,
"find_label_issues": find_issues,
}
@tool("inspect_dataset", return_direct=True)
def tool_inspect_dataset() -> Dict[str, Any]:
"""
Summarize the CURRENT dataset (no arguments required).
Behavior:
• Reads the dataset from the runtime context (set by the graph).
• Returns a compact summary of columns, dtypes, shape, and a guessed label/task.
Returns:
{
"type": "dataset_summary",
"columns": [...],
"dtypes": {col: dtype, ...},
"shape": (rows, cols),
"label_guess": "<name or None>",
"task_guess": "classification|regression|unsupervised",
"issues": [ ... ] # e.g., missing labels, single-class, etc.
}
"""
df_payload = get_df_payload() # default: current version
if df_payload is None:
raise RuntimeError("inspect_dataset: no dataset available in runtime context.")
df = df_from_payload(df_payload)
summary = guess_task_and_label(df)
# keep context fresh for downstream tools
set_df_summary(summary)
return {"type": "dataset_summary", **summary}
@tool("sota_preprocessing", return_direct=True)
def tool_sota_preprocessing(
task: Optional[str] = None,
modality: Optional[str] = None,
domain: Optional[str] = None,
target: Optional[str] = None,
) -> Dict[str, Any]:
"""
Search state-of-the-art preprocessing best practices (modality-aware).
Args:
task: e.g., "classification", "regression", "segmentation", "NER", "ASR", "forecasting".
If omitted, inferred from the dataset summary if available.
modality: one of {"tabular","text","image","audio","video","time_series","graph","multimodal"}.
domain: optional domain context (e.g., "clinical", "finance").
target: optional target structure (e.g., "segmentation masks", "bounding boxes").
Returns:
{
"type": "sota",
"task": ...,
"modality": ...,
"domain": ...,
"target": ...,
"queries": [...],
"bundled_results": [{ "query": q, "results": <tavily-results> }, ...],
"results": <first tavily-results batch>
}
"""
df_summary = get_df_summary() or {}
if not task:
task = df_summary.get("task_guess") or "classification"
yr = "2024 2025"
m = (modality or "").lower().strip()
modality_terms = {
"tabular": ["imputation", "encoding", "scaling", "outliers", "leakage prevention"],
"text": ["tokenization", "normalization", "subword", "BPE", "SentencePiece", "stopwords", "lemmatization", "augmentation"],
"image": ["normalization", "resizing", "color space", "augmentation", "RandAugment", "MixUp", "CutMix"],
"audio": ["resampling", "log-mel spectrogram", "MFCC", "pre-emphasis", "SpecAugment", "denoising"],
"time_series": ["resampling", "windowing", "detrending", "imputation", "outlier detection", "scaling"],
"video": ["frame sampling", "temporal augmentation", "clip normalization", "optical flow"],
"graph": ["feature normalization", "self-loops", "adjacency normalization", "sparsification"],
"multimodal": ["alignment", "synchronization", "fusion", "tokenization"],
}
m_terms = modality_terms.get(m, [])
# Build candidate queries
queries: List[str] = []
queries.append(f"state of the art preprocessing {task} {yr}")
queries.append(f"best practices data preprocessing {task} {yr}")
if m:
queries.append(f"{m} {task} preprocessing best practices {yr}")
if domain:
queries.append(f"{domain} {m or ''} {task} preprocessing best practices {yr}".strip())
if target:
queries.append(f"{m or ''} {task} {target} preprocessing pipeline {yr}".strip())
if m_terms:
queries.append(f"{m} {task} preprocessing {' '.join(m_terms)} {yr}")
# Deduplicate, preserve order
seen = set()
queries = [q for q in (q.strip() for q in queries) if q and (q not in seen and not seen.add(q))]
tavily = TavilySearchResults(k=6)
bundled: List[Dict[str, Any]] = [{"query": q, "results": tavily.invoke({"query": q})} for q in queries]
flat_first = bundled[0]["results"] if (bundled and "results" in bundled[0]) else []
# persist for planning
set_sota_bundled(bundled)
return {
"type": "sota",
"task": task,
"modality": m or "unknown",
"domain": domain,
"target": target,
"queries": queries,
"bundled_results": bundled,
"results": flat_first,
}
@tool("describe_step", return_direct=True)
def tool_describe_step(name: str) -> Dict[str, Any]:
"""
Return the exact docstring + parameter schema for a single step by name.
This prevents the model from inventing params.
"""
if name not in STEP_FUNCS:
raise ValueError(f"Unknown step '{name}'. Available: {list(STEP_FUNCS)}")
fn = STEP_FUNCS[name]
sig = get_signature_dict(fn) # your util that introspects defaults/annotations
return {"type": "step_description", "name": name, **sig}
@tool("list_steps", return_direct=True)
def tool_list_steps() -> Dict[str, Any]:
"""
List available pipeline steps (name, docstring, and signature).
Returns:
{
"type": "steps",
"steps": [
{
"name": "dedup" | "featurize" | "find_label_issues",
"doc": "<docstring>",
"params": [{"name": "...", "default": ..., "annotation": "...", "kind": "..."}]
}, ...
]
}
"""
return {
"type": "steps",
"steps": [{"name": n, **get_signature_dict(fn)} for n, fn in STEP_FUNCS.items()],
}
@tool("propose_plan", return_direct=True)
def tool_propose_plan(
task: Optional[str] = None,
modality: Optional[str] = None,
) -> Dict[str, Any]:
"""
Propose an ordered preprocessing plan grounded in SOTA + dataset summary.
(Planning only — does not execute steps.)
"""
df_summary = get_df_summary() or {}
bundled = get_sota_bundled() or []
if not task:
task = df_summary.get("task_guess") or "classification"
label_guess = df_summary.get("label_guess")
KEYWORDS = {
"dedup": {"duplicate", "near-duplicate", "near duplicate", "dupe", "dedup", "similarity", "knn", "kNN"},
"featurize": {
"impute", "imputation", "encoding", "one-hot", "scale", "scaling",
"normalize", "normalization", "standardize", "tfidf", "tokenization",
"lemmatization", "augmentation"
},
"find_label_issues": {"label noise", "noisy labels", "cleanlab", "confident learning", "label issues", "weak labels"},
}
def _score(text: str, keys: set[str]) -> int:
t = (text or "").lower()
return sum(1 for k in keys if k in t)
hits = {"dedup": 0, "featurize": 0, "find_label_issues": 0}
evidence: Dict[str, List[Dict[str, str]]] = {"dedup": [], "featurize": [], "find_label_issues": []}
for pack in bundled:
q = pack.get("query", "")
for item in (pack.get("results") or []):
title = item.get("title", "")
content = item.get("content", "")
url = item.get("url", "")
for step, keys in KEYWORDS.items():
s = _score(f"{q} {title} {content}", keys)
if s > 0:
hits[step] += s
if len(evidence[step]) < 5:
evidence[step].append({"query": q, "title": title, "url": url})
options: List[Dict[str, Any]] = []
if hits["dedup"] > 0 or modality in {None, "tabular", "text", "image", "time_series"}:
options.append(
{
"reason": "SOTA emphasizes handling near-duplicates early" if hits["dedup"] else "Practical first step to prevent leakage/skew",
"step": "dedup",
"params": {"threshold": 0.95, "metric": "cosine"},
"evidence": evidence["dedup"][:3],
}
)
options.append(
{
"reason": "SOTA emphasizes robust imputation/encoding/scaling" if hits["featurize"] else "Prepare features based on modality",
"step": "featurize",
"params": {"nan_strategy": "impute"},
"evidence": evidence["featurize"][:3],
}
)
if (task == "classification" and label_guess) or hits["find_label_issues"] > 0:
options.append(
{
"reason": "SOTA recommends checking noisy labels" if hits["find_label_issues"] else "Check label quality before training",
"step": "find_label_issues",
"params": {"label": label_guess or "<CONFIRM>"},
"evidence": evidence["find_label_issues"][:3],
}
)
if not options:
options = [{"reason": "Generic best practice", "step": "featurize", "params": {"nan_strategy": "impute"}, "evidence": []}]
return {
"type": "plan",
"task": task,
"modality": modality,
"label_guess": label_guess,
"options": options,
"keyword_hits": hits,
}
@tool("run_step", return_direct=True)
def tool_run_step(name: str, params_json: str = "") -> Dict[str, Any]:
"""
Execute a single pipeline step on the CURRENT dataset (no df argument).
Returns ONLY a compact summary; the updated df is stored in runtime context.
"""
df_payload = get_df_payload()
if df_payload is None:
raise RuntimeError("run_step: no dataset available in runtime context.")
if name not in STEP_FUNCS:
raise ValueError(f"Unknown step '{name}'. Available: {list(STEP_FUNCS)}")
params = json.loads(params_json) if params_json else {}
if not isinstance(params, dict):
raise ValueError("params_json must decode to a JSON object")
df = df_from_payload(df_payload)
df_out, stats = STEP_FUNCS[name](df=df, **params)
df_next = df_out if df_out is not None else df
# ✅ update runtime dataset, but DO NOT send it back in the tool message
set_df_payload(df_to_payload(df_next))
# Build a tiny, safe summary for the model
shape_before = (len(df), len(df.columns))
shape_after = (len(df_next), len(df_next.columns))
compact_stats = {k: stats.get(k) for k in [
"n_rows_before_dedup", "n_near_dupe_pairs", "n_groups",
"n_rows_flagged_duplicates", "n_rows_after_dedup",
"metric", "threshold", "k", "total_time_sec"
] if k in stats}
return {
"type": "step_result",
"name": name,
"params_used": params,
"shape_before": shape_before,
"shape_after": shape_after,
"stats": compact_stats, # small dict only
"note": "Dataset updated in runtime context; use list_versions/reset_to_version if needed."
}
# ---------------------------
# Optional helpers for version control from chat/agent
# ---------------------------
@tool("list_versions", return_direct=True)
def tool_list_versions() -> Dict[str, Any]:
"""
Return a lightweight view of the version stack:
{ count, current_index, has_base, meta: [{...}, ...] }
"""
return {"type": "versions", **_list_versions_state()}
@tool("reset_to_version", return_direct=True)
def tool_reset_to_version(spec: str) -> Dict[str, Any]:
"""
Move CURRENT pointer to a prior version without deleting history.
spec can be: "base" | "prev" | "@-1" | "@-2" | "3"
"""
# accept int or @-k in string form
try:
if spec.isdigit():
_reset_current_to(int(spec))
else:
_reset_current_to(spec)
except Exception as e:
raise RuntimeError(f"reset_to_version: {e}")
df_payload = get_df_payload()
return {
"type": "reset",
"current": _list_versions_state(),
"df": df_payload,
}
|