webfetch-llm 0.1.0__py3-none-any.whl

webfetch/compress.py

@@ -0,0 +1,360 @@
+"""
+Sentence-level extractive compression for tool results.
+
+Within each top-ranked chunk, keep only the sentences relevant to the query.
+Runs at tool-result FORMATTING time, after the cache read, so cached chunks
+stay full-text and budget-agnostic - retuning compression never invalidates
+the cache.
+
+Why extractive (select sentences) rather than abstractive (rewrite):
+abstractive needs an LLM call per result, which is against the library's
+zero-marginal-cost design; extractive is a pure function of (query, chunks).
+
+Parameters (scorer, selection policy, guards) are chosen by the eval harness
+(evals/run_compression_eval.py) which imports THIS module - the shipped code
+is exactly what was measured. Config defaults in webfetch/config.py hold the
+winning configuration: cross-encoder scored, ratio 0.5, all guards on -
+measured ZERO recall drop (29/29 answer survival) at 50% of baseline tokens
+on 50 captured production results. The cross-encoder is what makes lossless
+compression possible here: bi-encoder selection at the same token level
+loses 3-4 answers of 29.
+
+Guards against the known failure modes of naive sentence selection:
+  - anaphora: a kept sentence starting with a pronoun/connective would be
+    unreadable alone, so its preceding sentence is retained with it
+  - table-like lines (markdown rows, digit-dense spec lines) are not prose -
+    sentence scoring misjudges them, so they bypass scoring and are kept
+  - cross-chunk dedup: the chunker's 10% overlap re-emits boundary sentences
+    in the next chunk; duplicates waste budget
+
+Requires the optional rerank extra for the bi-encoder scorer; without it the
+scorer degrades to the zero-dep lexical scorer (never crashes the tool path).
+"""
+
+from __future__ import annotations
+
+import logging
+import math
+import re
+import threading
+
+from webfetch.config import (
+    BIENCODER_MODEL,
+    COMPRESS_ANAPHORA_GUARD,
+    COMPRESS_DEDUP,
+    COMPRESS_PARAM,
+    COMPRESS_POLICY,
+    COMPRESS_SCORER,
+    COMPRESS_TABLE_GUARD,
+    CROSSENCODER_MODEL,
+)
+from webfetch.rank.base import Chunk
+
+logger = logging.getLogger(__name__)
+
+# Sentence boundary: terminal punctuation, whitespace, then something that
+# looks like a sentence start. Deliberately regex-based and zero-dep (same
+# philosophy as the chunker) - abbreviation mistakes cost a split, and the
+# eval measures the net effect rather than assuming splitter perfection.
+_SENT_BOUNDARY_RE = re.compile(r"(?<=[.!?])\s+(?=[A-Z0-9\"'(\[])")
+
+# First words that signal the sentence leans on its predecessor.
+_ANAPHORIC_STARTS = frozenset({
+    "it", "its", "this", "that", "these", "those", "they", "them", "their",
+    "he", "she", "his", "her", "him", "however", "but", "also", "moreover",
+    "additionally", "furthermore", "meanwhile", "instead", "still", "yet",
+    "so", "such", "both", "neither", "otherwise", "consequently", "thus",
+    "therefore", "here", "there",
+})
+
+_FIRST_WORD_RE = re.compile(r"[a-zA-Z]+")
+
+# Digit share of non-space chars above which a line is treated as data,
+# not prose (spec lines, score lines, table rows).
+_TABLE_DIGIT_RATIO = 0.20
+
+# Sentences shorter than this are never deduped - short strings ("Yes.",
+# a bare number) can legitimately repeat with different meanings.
+_DEDUP_MIN_CHARS = 30
+
+_bi_model = None
+_ce_model = None
+_models_unavailable = False
+# Concurrent tool calls racing lazy model construction crash inside torch -
+# same locking pattern as rank/biencoder.py.
+_init_lock = threading.Lock()
+
+
+def split_sentences(text: str) -> list[tuple[int, str]]:
+    """Split text into (line_index, sentence) pairs.
+
+    Newlines are hard boundaries (extracted pages use them structurally:
+    headings, list items, table rows); prose lines are further split on
+    sentence punctuation. Chunk-edge fragments come through as ordinary
+    sentences and are scored like any other.
+
+    Args:
+        text: Chunk text.
+
+    Returns:
+        List of (line_index, sentence) pairs in document order.
+    """
+    out: list[tuple[int, str]] = []
+    for line_idx, line in enumerate(text.split("\n")):
+        line = line.strip()
+        if not line:
+            continue
+        for sent in _SENT_BOUNDARY_RE.split(line):
+            sent = sent.strip()
+            if sent:
+                out.append((line_idx, sent))
+    return out
+
+
+def is_table_like(sentence: str) -> bool:
+    """True for lines that are data rather than prose (skip scoring)."""
+    if "|" in sentence:
+        return True
+    compact = sentence.replace(" ", "")
+    if not compact:
+        return False
+    digits = sum(ch.isdigit() for ch in compact)
+    return digits / len(compact) >= _TABLE_DIGIT_RATIO
+
+
+def starts_anaphoric(sentence: str) -> bool:
+    """True if the sentence's first word leans on the preceding sentence."""
+    m = _FIRST_WORD_RE.search(sentence[:24])
+    return bool(m) and m.group(0).lower() in _ANAPHORIC_STARTS
+
+
+def _tokenize(text: str) -> list[str]:
+    return re.sub(r"[^\w\s]", " ", text.lower()).split()
+
+
+def score_lexical(query: str, sentences: list[str]) -> list[float]:
+    """IDF-weighted query-term coverage per sentence, in [0, 1].
+
+    IDF is computed over the sentences of THIS call (the per-result corpus):
+    a query term appearing in every sentence carries no selection signal,
+    while a rare one dominates. Zero-dep fallback scorer.
+    """
+    q_terms = list(dict.fromkeys(_tokenize(query)))
+    if not q_terms or not sentences:
+        return [0.0] * len(sentences)
+    sent_tokens = [set(_tokenize(s)) for s in sentences]
+    n = len(sentences)
+    idf = {t: math.log(1 + n / (1 + sum(t in st for st in sent_tokens)))
+           for t in q_terms}
+    total = sum(idf.values()) or 1.0
+    return [sum(idf[t] for t in q_terms if t in st) / total
+            for st in sent_tokens]
+
+
+def _deps_available() -> bool:
+    global _models_unavailable
+    if _models_unavailable:
+        return False
+    try:
+        import sentence_transformers  # noqa: F401
+    except ImportError:
+        logger.warning(
+            "sentence-transformers not installed - compression degrades to "
+            "the lexical scorer. Install webfetch-llm[rerank]."
+        )
+        _models_unavailable = True
+        return False
+    return True
+
+
+def score_biencoder(query: str, sentences: list[str]) -> list[float] | None:
+    """Cosine similarity of each sentence to the query; None without [rerank]."""
+    global _bi_model
+    if not _deps_available():
+        return None
+    from sentence_transformers import SentenceTransformer
+    with _init_lock:
+        if _bi_model is None:
+            _bi_model = SentenceTransformer(BIENCODER_MODEL)
+    q_vec = _bi_model.encode([query], normalize_embeddings=True,
+                             show_progress_bar=False)[0]
+    s_vecs = _bi_model.encode(sentences, normalize_embeddings=True,
+                              show_progress_bar=False)
+    return [float(v @ q_vec) for v in s_vecs]
+
+
+def score_crossencoder(query: str, sentences: list[str]) -> list[float] | None:
+    """Joint (query, sentence) relevance scores; None without [rerank].
+
+    Same ms-marco model as the ranking cascade's stage 3. O(N) forward
+    passes, but N here is the ~15-25 sentences of one tool result - tens of
+    milliseconds, not the cost profile that keeps cross-encoders out of
+    full-list ranking.
+    """
+    global _ce_model
+    if not _deps_available():
+        return None
+    from sentence_transformers import CrossEncoder
+    with _init_lock:
+        if _ce_model is None:
+            _ce_model = CrossEncoder(CROSSENCODER_MODEL)
+    scores = _ce_model.predict([(query, s) for s in sentences],
+                               show_progress_bar=False)
+    return [float(s) for s in scores]
+
+
+def _select_indices(scores: list[float], sentences: list[str],
+                    policy: str, param: float) -> set[int]:
+    """Pick sentence indices for one chunk under the selection policy.
+
+    Policies:
+        ratio: best-scoring sentences until ~param of the chunk's chars kept.
+        topk: best int(param) sentences.
+        threshold: every sentence scoring >= param.
+
+    Always returns at least one index (the best-scoring sentence) so a chunk
+    can never compress to nothing.
+    """
+    order = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)
+    keep: set[int] = set()
+    if policy == "ratio":
+        total = sum(len(s) for s in sentences)
+        used = 0
+        for i in order:
+            if keep and used + len(sentences[i]) > param * total:
+                continue  # try smaller later sentences - greedy knapsack
+            keep.add(i)
+            used += len(sentences[i])
+    elif policy == "topk":
+        keep = set(order[: max(1, int(param))])
+    elif policy == "threshold":
+        keep = {i for i in order if scores[i] >= param}
+    else:
+        raise ValueError(f"unknown compression policy: {policy!r}")
+    if not keep:
+        keep = {order[0]}
+    return keep
+
+
+def compress_chunks(
+    query: str,
+    chunks: list[Chunk],
+    scorer: str = COMPRESS_SCORER,
+    policy: str = COMPRESS_POLICY,
+    param: float = COMPRESS_PARAM,
+    anaphora_guard: bool = COMPRESS_ANAPHORA_GUARD,
+    table_guard: bool = COMPRESS_TABLE_GUARD,
+    dedup: bool = COMPRESS_DEDUP,
+) -> list[Chunk]:
+    """Compress each chunk to its query-relevant sentences.
+
+    Returns NEW Chunk objects - inputs are never mutated, because cached
+    chunk instances may be shared with other readers. Kept sentences stay in
+    document order; sentences from the same source line rejoin with a space,
+    different lines with a newline (preserves table/list structure).
+
+    On any unexpected failure the original chunks are returned unchanged -
+    this runs on the tool path, where a formatting bug must never turn a
+    good search into an error.
+
+    Args:
+        query: The search query driving relevance.
+        chunks: Ranked chunks (best first) from the pipeline.
+        scorer: "biencoder", "crossencoder", "lexical", or "lead"
+            (position baseline).
+        policy: Selection policy - "ratio", "topk", or "threshold".
+        param: Policy parameter (ratio fraction, k, or score cutoff).
+        anaphora_guard: Retain the predecessor of pronoun-initial keeps.
+        table_guard: Keep table-like lines unconditionally.
+        dedup: Drop sentences already emitted by an earlier chunk.
+
+    Returns:
+        Compressed chunks, same order; chunks fully deduplicated away are
+        dropped. Empty input returns [].
+    """
+    if not chunks:
+        return []
+    try:
+        return _compress(query, chunks, scorer, policy, param,
+                         anaphora_guard, table_guard, dedup)
+    except Exception:
+        logger.warning("compression failed - returning chunks uncompressed",
+                       exc_info=True)
+        return chunks
+
+
+def _compress(query: str, chunks: list[Chunk], scorer: str, policy: str,
+              param: float, anaphora_guard: bool, table_guard: bool,
+              dedup: bool) -> list[Chunk]:
+    per_chunk = [split_sentences(c.text) for c in chunks]
+
+    # Score every sentence across all chunks in ONE batch (matters for the
+    # bi-encoder: one encode call instead of one per chunk).
+    flat = [s for sents in per_chunk for (_, s) in sents]
+    if scorer == "biencoder":
+        flat_scores = score_biencoder(query, flat)
+        if flat_scores is None:
+            flat_scores = score_lexical(query, flat)
+    elif scorer == "crossencoder":
+        flat_scores = score_crossencoder(query, flat)
+        if flat_scores is None:
+            flat_scores = score_lexical(query, flat)
+    elif scorer == "lexical":
+        flat_scores = score_lexical(query, flat)
+    elif scorer == "lead":
+        # Position baseline: earlier in chunk = higher. Eval control arm.
+        flat_scores = []
+        for sents in per_chunk:
+            flat_scores.extend(1.0 / (1 + i) for i in range(len(sents)))
+    else:
+        raise ValueError(f"unknown compression scorer: {scorer!r}")
+
+    out: list[Chunk] = []
+    seen_norm: set[str] = set()
+    pos = 0
+    for chunk, sents in zip(chunks, per_chunk):
+        scores = flat_scores[pos: pos + len(sents)]
+        pos += len(sents)
+        if not sents:
+            continue
+        if len(sents) == 1:
+            keep = {0}
+        else:
+            texts = [s for (_, s) in sents]
+            keep = _select_indices(scores, texts, policy, param)
+            if table_guard:
+                keep |= {i for i, t in enumerate(texts) if is_table_like(t)}
+            if anaphora_guard:
+                # Walk each kept sentence's predecessors so chains resolve
+                # ("It ... " preceded by "This ..." pulls in both).
+                for i in sorted(keep):
+                    j = i
+                    while (j > 0 and starts_anaphoric(texts[j])
+                           and (j - 1) not in keep):
+                        keep.add(j - 1)
+                        j -= 1
+
+        kept = [sents[i] for i in sorted(keep)]
+        if dedup:
+            deduped = []
+            for line_idx, s in kept:
+                norm = " ".join(_tokenize(s))
+                if len(s) >= _DEDUP_MIN_CHARS and norm in seen_norm:
+                    continue
+                seen_norm.add(norm)
+                deduped.append((line_idx, s))
+            kept = deduped
+        if not kept:
+            continue  # chunk was pure overlap-duplicate content
+
+        parts = [kept[0][1]]
+        for (prev_line, _), (line_idx, s) in zip(kept, kept[1:]):
+            parts.append(("\n" if line_idx != prev_line else " ") + s)
+        out.append(Chunk(text="".join(parts), url=chunk.url,
+                         title=chunk.title, score=chunk.score))
+    return out
+
+
+__all__ = ["compress_chunks", "split_sentences", "score_lexical",
+           "score_biencoder", "score_crossencoder", "is_table_like",
+           "starts_anaphoric"]

webfetch/config.py

@@ -0,0 +1,154 @@
+"""
+Central constants and defaults for the webfetch pipeline.
+
+All tuneable knobs live here so nothing is hardcoded in module logic.
+Change a value here and it propagates everywhere - no grep-and-replace needed.
+"""
+
+# --- Search ---
+# How many URLs to retrieve from the search adapter per query.
+DEFAULT_N_RESULTS: int = 10
+
+# Which search provider to use when none is specified by the caller.
+# "ddg" = DuckDuckGo (free, no key); see search/__init__.py for all.
+DEFAULT_SEARCH_PROVIDER: str = "ddg"
+
+# --- Fetch ---
+# Seconds before an HTTP fetch is abandoned. Keeps slow pages from
+# blocking the pipeline.
+FETCH_TIMEOUT_SECS: int = 10
+
+# Thread pool size for concurrent URL fetching. Fetching is IO-bound (network
+# waits, not CPU), so more workers than cores is fine and cuts wall time a lot.
+DEFAULT_FETCH_WORKERS: int = 8
+
+# Extracted text shorter than this suggests a JS-rendered SPA whose real
+# content never reached the prose extractor - triggers the playwright
+# fallback even though extraction technically "succeeded".
+MIN_EXTRACTED_CHARS: int = 300
+
+# --- Ranking ---
+# Default cascade: hybrid fusion (full-list BM25 + bi-encoder rankings fused
+# via RRF) -> cross-encoder. Fusion replaced the BM25-first gate after the
+# gap-1 experiment measured recall 46% -> 54% at identical token cost - the
+# lexical gate was dropping semantically-relevant chunks it could not see.
+HYBRID_FUSION_TOP_K: int = 30  # fused chunks passed to the cross-encoder
+
+# Used by the degraded (no-embeddings) path and by callers composing the
+# old-style cascade manually.
+BM25_TOP_K: int = 20       # after BM25: keep best 20 chunks
+BIENCODER_TOP_K: int = 10  # after bi-encoder: keep best 10
+CROSSENCODER_TOP_K: int = 5  # after cross-encoder: keep best 5
+
+# Default models for optional reranking stages.
+BIENCODER_MODEL: str = "all-MiniLM-L6-v2"
+CROSSENCODER_MODEL: str = "cross-encoder/ms-marco-MiniLM-L-6-v2"
+
+# --- Chunking ---
+# Characters per text chunk fed into the ranker.
+# 400 chars ~ 100 tokens - small enough for precise reranking, big
+# enough for context.
+DEFAULT_CHUNK_SIZE: int = 400
+
+# Overlap between consecutive chunks to avoid splitting a key sentence
+# at a boundary.
+CHUNK_OVERLAP_RATIO: float = 0.10  # 10% overlap
+
+# --- Token budget ---
+# Maximum characters of ranked text sent to the LLM extraction call.
+# ~6000 chars / 4 chars-per-token ~ 1500 tokens of context - cheap for
+# haiku-class models.
+DEFAULT_TOKEN_BUDGET: int = 6000
+
+# --- Cache ---
+# Default path for the sqlite3 cache database.
+DEFAULT_CACHE_DB: str = "~/.webfetch/cache.db"
+
+# How long cached results are kept before expiry. Acts as the hard ceiling
+# for all rows (pages, legacy query rows without a freshness class).
+DEFAULT_CACHE_TTL_DAYS: int = 90
+
+# Volatility-aware TTLs, resolved at READ time so retuning these applies to
+# already-cached rows. Effective TTL = min(TTL of the row's stored class,
+# TTL of the caller's freshness hint, the ttl_days ceiling).
+TTL_BY_FRESHNESS: dict[str, int] = {
+    "realtime": 15 * 60,      # prices, scores, breaking news
+    "recent": 7 * 86400,      # things that change over weeks/months
+    "stable": 90 * 86400,     # historical/definitional facts, specs
+}
+
+# Class used when there is no model hint and the classifier is unsure,
+# and for legacy rows cached before freshness existed.
+DEFAULT_FRESHNESS: str = "recent"
+
+# --- Search resilience ---
+# Circuit breaker: an engine failing this many CONSECUTIVE times (errors,
+# or silent-block empties) is benched for the cooldown, then given one
+# half-open probe. Stops burning latency on a dead/blocking engine.
+SEARCH_BREAKER_THRESHOLD: int = 3
+SEARCH_BREAKER_COOLDOWN_SECS: float = 300.0
+
+# In fusion, an engine's empty response only counts as a breaker failure
+# when a peer engine returned at least this many results for the SAME
+# query - distinguishes a silent block from a genuinely hard query.
+FUSION_PEER_EMPTY_MIN: int = 3
+
+# --- Cost receipts ---
+# Baseline for savings estimates (webfetch/receipts.py). Both Anthropic and
+# OpenAI charge $10/1k hosted searches (verified 2026-07); the hosted tools
+# also inject ~17.4k content tokens per call vs our ~3.5k compressed
+# (measured, Layer 3 eval 2026-07-14). Token price defaults to Opus-class
+# input - override per call for cheaper models.
+RECEIPT_HOSTED_SEARCH_FEE: float = 10.00 / 1000
+RECEIPT_HOSTED_TOKENS_PER_CALL: int = 17400
+RECEIPT_TOKEN_PRICE_PER_MTOK: float = 5.00
+
+# --- LLM extraction ---
+# Default model for the extraction step. Cheapest capable model by default.
+DEFAULT_EXTRACT_MODEL: str = "claude-haiku-4-5-20251001"
+
+# --- Semantic query cache ---
+# Thresholds picked by the eval harness (evals/run_matcher_eval.py, 247-pair
+# factoid-enriched set): bi >= 0.60 keeps paraphrase recall at 0.98 as a
+# shortlist gate; NLI bidirectional entailment >= 0.97 gives precision 0.955
+# with zero trusted-negative false positives. See matcher_recommendation.json.
+SEMCACHE_BI_THRESHOLD: float = 0.60
+SEMCACHE_CE_THRESHOLD: float = 0.97
+
+# NLI verifier won the 4-model bake-off: only model to clear the 0.95
+# precision bar, natively rejects entity/number swaps (Portland trap: 0.000).
+SEMCACHE_CE_MODEL: str = "cross-encoder/nli-deberta-v3-base"
+
+# Verify only the top cosine candidate - precision-first; Layer 2 measured
+# zero wrong-target matches, so deeper candidate lists buy nothing.
+SEMCACHE_MAX_CANDIDATES: int = 1
+
+# --- Tool mode ---
+# Max characters of ranked context returned per web_search tool call.
+# Larger than DEFAULT_TOKEN_BUDGET because in tool mode the calling model is
+# the extractor and benefits from a bit more surrounding context.
+DEFAULT_TOOL_RESULT_BUDGET: int = 8000
+
+# --- Sentence-level compression (tool-result formatting) ---
+# Applied AFTER the cache read, so cached rows stay full-text and retuning
+# these never invalidates the cache. Values picked by the eval sweep
+# (evals/run_compression_eval.py, 50 captured production results): this
+# config measured recall 29/50 vs 29/50 uncompressed (zero drop, 29/29
+# answer survival) at 50% of baseline tokens (332 vs 665 mean). The
+# cross-encoder scorer is what preserves recall - biencoder at the same
+# tokens loses 3-4 answers. Without [rerank] it degrades to the lexical
+# scorer (measured: 27/29 survival at 51%).
+COMPRESSION_ENABLED: bool = True
+COMPRESS_SCORER: str = "crossencoder"  # | "biencoder" | "lexical" | "lead"
+COMPRESS_POLICY: str = "ratio"         # "ratio" | "topk" | "threshold"
+COMPRESS_PARAM: float = 0.5     # keep best sentences, <= 50% of chunk chars
+COMPRESS_ANAPHORA_GUARD: bool = True  # +14 tokens, keeps pronouns resolvable
+COMPRESS_TABLE_GUARD: bool = True      # ablation: protects 1 answer of 29
+COMPRESS_DEDUP: bool = True     # ablation: -13 tokens free (overlap dupes)
+
+# Tool-result context format (build_context). Same eval: merging same-URL
+# chunks under one header + hostname-only headers cut the fixed header
+# overhead (26% of an uncompressed result) with zero measured recall cost -
+# titles stay because answers live in them (extraction-hardening eval).
+TOOL_MERGE_SOURCES: bool = True
+TOOL_HEADER_STYLE: str = "domain"      # "full" | "domain"

webfetch/data/volatility_centroids.json

@@ -0,0 +1 @@
+{"model": "all-MiniLM-L6-v2", "source": "FreshQA (CC), full 377 rows", "classes": {"realtime": [0.012145215645432472, -0.0018631262937560678, -0.06274841725826263, -0.007661583833396435, -0.027888312935829163, 0.0621718093752861, -0.021171551197767258, 0.02851240523159504, -0.011449315585196018, 0.02597236819565296, -0.04086684435606003, -0.006424089893698692, -0.02788771130144596, 0.02362537570297718, -0.03769982233643532, 0.017427198588848114, -0.017687330022454262, -0.08234846591949463, 0.08868501335382462, -0.10261496156454086, -0.007088787388056517, 0.002663521096110344, 0.0016332603991031647, 0.00045282961218617857, 0.08852767199277878, -0.03599361702799797, -0.056443117558956146, -0.022277243435382843, -0.07022085040807724, -0.016778215765953064, -0.06464487314224243, -0.02756372094154358, 0.027233166620135307, 0.03248134255409241, -0.04783632978796959, -0.04881690815091133, 0.02403809316456318, 0.0007506250985898077, 0.02884712442755699, 0.012483729049563408, -0.02234613709151745, -0.10250423848628998, 0.016286777332425117, 0.012731242924928665, 0.03766198456287384, -0.025340810418128967, -0.016598522663116455, 0.02768656052649021, 0.054936714470386505, 0.11487270146608353, -0.03534339368343353, 0.03725761920213699, -0.017243774607777596, 0.009950782172381878, 0.06690368056297302, 0.005692985374480486, -0.044698603451251984, 0.0004932482843287289, 0.04355696961283684, 0.01979084312915802, -0.007639622315764427, -0.03338771313428879, -0.10322568565607071, 0.03135615959763527, 0.01046554371714592, -0.04580345377326012, -0.026506880298256874, -0.07263809442520142, -0.014567894861102104, -0.02869809791445732, 0.01659376546740532, 0.010415657423436642, 0.006484807003289461, -0.08891996741294861, 0.023524625226855278, 0.01739215850830078, 0.035195961594581604, 0.03747803717851639, 0.003209460061043501, -0.0376424640417099, 0.007335755508393049, -0.08569755405187607, -0.015708794817328453, -0.021274028345942497, 0.056080762296915054, 0.03369034081697464, -0.032607536762952805, 0.07464656978845596, -0.056957997381687164, 0.014871630817651749, -0.05218343064188957, 0.029951659962534904, -0.01689605414867401, 0.052715156227350235, -0.1119815930724144, -0.0018735149642452598, -0.028360188007354736, -0.004249660298228264, -0.04698532447218895, 0.11000624299049377, -0.021992405876517296, 0.07617287337779999, 0.042818401008844376, 0.0037722475826740265, 0.056841593235731125, 0.030398523434996605, 0.06313391029834747, 0.028064215555787086, -0.033619530498981476, -0.026942992582917213, -0.05855797603726387, -0.009410723112523556, -0.023611513897776604, -0.0013498563785105944, -0.004817285109311342, -0.04733685776591301, 0.04180905967950821, 0.0655769407749176, 0.06571927666664124, 0.029141250997781754, 0.06606484204530716, 0.04993269219994545, 0.0006714511546306312, 0.010644162073731422, -0.09988965839147568, 0.016641344875097275, -0.017963502556085587, -7.996425952427355e-33, 0.04056042432785034, -0.08298078179359436, 0.0798611044883728, 0.03384505212306976, -0.0005813181633129716, -0.013754554092884064, 0.008965825662016869, -0.018057813867926598, -0.055044565349817276, -0.07378089427947998, -0.019387418404221535, -0.06616166234016418, -0.041721466928720474, -0.11676525324583054, 0.0884120985865593, -0.0005623944452963769, -0.08393930643796921, 0.045821357518434525, -0.03545466437935829, 0.02201877534389496, 0.02102014608681202, 0.03818574175238609, 0.005251807626336813, -0.046063873916864395, 0.0017609883798286319, -0.012767121195793152, 0.016234265640378, -0.03738025575876236, 0.03839366137981415, 0.017702419310808182, 0.02254652976989746, -0.011753003112971783, -0.008132238872349262, 0.021470986306667328, 0.026390258222818375, 0.031572166830301285, -0.02854541316628456, -0.04545816779136658, 0.05124926194548607, 0.032728858292102814, 0.014105470851063728, -0.03141537681221962, 0.005319760646671057, -0.0634508728981018, -0.012997179292142391, 0.016193777322769165, -0.0289461612701416, 0.013243574649095535, 0.07796746492385864, 0.019404128193855286, -0.018053743988275528, -0.00757976621389389, -0.08551377803087234, -0.017113754525780678, 0.01956339366734028, -0.0217848252505064, 0.028527241200208664, -0.052100133150815964, 0.08539976924657822, 0.0727059468626976, -0.00771735142916441, 0.10828250646591187, -0.0010062677320092916, -0.0017656820127740502, -0.07601210474967957, 0.03872839733958244, 0.06104473024606705, -0.016915468499064445, -0.054150838404893875, 0.04271630197763443, 0.07039860635995865, -0.0031229224987328053, 0.010230539366602898, -0.0991160199046135, 0.017362629994750023, 0.013642113655805588, 0.0605914480984211, -0.0418471023440361, 0.02950374223291874, 0.060241036117076874, -0.047129079699516296, 0.007117462810128927, 0.07195363193750381, 0.028437813743948936, -0.06897237151861191, 0.02679777517914772, -0.042234260588884354, 0.026559779420495033, 0.012453213334083557, -0.025323128327727318, -0.062185730785131454, -0.07120354473590851, 0.03637091815471649, -0.04062645882368088, -0.06572014838457108, 3.756538684524308e-33, -0.019561057910323143, 0.0010238690301775932, 0.03263934329152107, 0.06091630086302757, 0.03326853737235069, -0.05022892355918884, 0.02662106789648533, 0.20079641044139862, 0.02190878428518772, 0.03251644968986511, 0.10060176253318787, -0.005975282751023769, 0.11721522361040115, 0.03066098317503929, 0.07437632977962494, 0.054939668625593185, 0.058271657675504684, -0.047429993748664856, -0.06029954552650452, -0.03828994557261467, -0.024580322206020355, 0.0824195146560669, -0.05107038468122482, 0.029964735731482506, 0.02194763906300068, -0.013980412855744362, 0.053490251302719116, 0.00391295924782753, -0.13898153603076935, -0.08299368619918823, 0.007824414409697056, -0.012302936986088753, -0.1006881594657898, 0.04606233164668083, -0.03258052095770836, 0.09304725378751755, 0.006648424081504345, -0.00026814319426193833, 0.028638750314712524, 0.07307504117488861, 0.05531078577041626, -0.017754772678017616, -0.012831241823732853, 0.16821259260177612, 0.05074983090162277, 0.03985459730029106, -0.05292412266135216, 0.08841066062450409, 0.05655524879693985, -0.008059504441916943, -0.07679089158773422, 0.026545094326138496, -0.07387138158082962, 0.01698480173945427, 0.01564045250415802, 0.03352392092347145, -0.07063072174787521, 0.03391892462968826, -0.013085504062473774, -0.018996283411979675, 0.008864144794642925, 0.026843387633562088, -0.04298478364944458, -0.009842624887824059, -0.013877437449991703, 7.742519665043801e-05, 0.01939546875655651, 0.03087269701063633, -0.06924799829721451, -0.010178822092711926, 0.05894096940755844, 0.006799302063882351, -0.06124420464038849, -0.021927375346422195, -0.09395230561494827, 0.10161103308200836, 0.05908716842532158, 0.04067505523562431, 0.03260904178023338, -0.020806023851037025, -0.0950707197189331, 0.053077954798936844, -0.056930337101221085, 0.044653572142124176, -0.012820999138057232, -0.003900255775079131, 0.06577984988689423, -0.1083097755908966, -0.018936049193143845, 0.04814876988530159, -0.03550277650356293, -0.0239479448646307, -0.004242108669131994, -0.09287764877080917, -0.017485903576016426, -4.011367948919542e-08, 0.05853461101651192, 0.14321114122867584, -0.09642613679170609, -0.068077951669693, 0.05938224494457245, 0.06424376368522644, 0.06816771626472473, 0.02329518087208271, 0.024788212031126022, 0.0216663908213377, 0.10322057455778122, -0.04315197840332985, 0.01926328055560589, -0.01207323744893074, -0.0005831048474647105, -0.04060975834727287, -0.025956114754080772, 0.04761594533920288, -0.008457469753921032, -0.041826892644166946, 0.026088401675224304, 0.031063469126820564, 0.07961023598909378, -0.09879054129123688, 0.0010286689503118396, 0.004794402047991753, -0.013164789415895939, 0.05569709464907646, -0.015505952760577202, -0.005806891713291407, 0.019580019637942314, 0.016695119440555573, -0.08401018381118774, -0.14084604382514954, -0.024840427562594414, 0.0046034264378249645, -0.04590464383363724, 0.02744842879474163, -0.03233516216278076, -0.08428578823804855, 0.010429917834699154, 0.06057283282279968, 0.010982648469507694, 0.06134016811847687, 0.04124429449439049, 0.029994413256645203, -0.023568596690893173, -0.037896428257226944, 0.020653579384088516, -0.08995361626148224, -0.07378672063350677, 0.02678094245493412, 0.054825007915496826, -0.020050644874572754, 0.019060198217630386, 0.002838355489075184, -0.025905828922986984, -0.046560149639844894, -0.06630369275808334, -0.03418434411287308, 0.11502895504236221, -0.12107022851705551, 0.04983608052134514, 0.08930587023496628], "recent": [0.037740834057331085, -0.025897791609168053, -0.07063736766576767, -0.0030858858954161406, -0.0385655052959919, 0.027679670602083206, -0.005790873430669308, 0.04536600410938263, 0.009580287151038647, 0.041824787855148315, 0.002868938259780407, -0.0330912210047245, 0.010563128627836704, 0.024402238428592682, -0.03492194414138794, 0.01100903283804655, -0.04160817340016365, -0.07459341734647751, 0.0861305445432663, -0.11681894958019257, 0.000607187335845083, -0.025019336491823196, 0.04239147529006004, -0.030158009380102158, 0.07918523997068405, -0.026850827038288116, -0.06338309496641159, -0.008769571781158447, -0.05210849270224571, -0.0670437440276146, -0.05640832707285881, -0.03528883308172226, 0.04560999572277069, 0.04130649194121361, -0.011238071136176586, -0.028030933812260628, 0.025195254012942314, 0.008335919119417667, 0.06000549718737602, -0.007748901378363371, 0.00048418165533803403, -0.07644284516572952, 0.03349863365292549, -0.007790022064000368, 0.0558280423283577, -0.0007756453705951571, -0.0057226321659982204, 0.02470877580344677, 0.020756904035806656, 0.07881532609462738, -0.005602407269179821, 0.02508022077381611, -0.016782615333795547, 0.03592527285218239, 0.07109124958515167, -0.04779466986656189, -0.04691481590270996, -0.0009210046846419573, 0.025139687582850456, 0.053828924894332886, -0.002145996317267418, -0.017301000654697418, -0.11765002459287643, 0.01028557401150465, -0.011623680591583252, -0.020329630002379417, -0.033269159495830536, -0.0458010695874691, -0.03787567839026451, 0.0015074169496074319, 0.05620766058564186, -0.018920497968792915, 0.010189317166805267, -0.03809700906276703, -0.0008899683598428965, -0.01289956271648407, 0.05275963246822357, 0.09183822572231293, -0.0025698968674987555, -0.05026788264513016, 0.03196341544389725, -0.04640628397464752, -0.02241269126534462, -0.01151073258370161, 0.05161017179489136, 0.0334722064435482, -0.028133969753980637, 0.060756903141736984, -0.05072283744812012, -0.0039917235262691975, -0.0464647077023983, 0.04364056885242462, 0.02522849477827549, 0.0595581941306591, -0.10544005036354065, 0.040252745151519775, -0.035451192408800125, 0.03506895899772644, -0.06946254521608353, 0.09570952504873276, -0.015216988511383533, 0.049334000796079636, 0.07752926647663116, 0.013933182694017887, -0.01580444537103176, 0.03478824347257614, 0.04229752719402313, 0.07602357864379883, -0.015588724054396152, 0.002205786295235157, -0.07989031821489334, -0.008642822504043579, -0.05011498183012009, -0.005334169138222933, -0.005515052936971188, -0.05893829092383385, 0.060996122658252716, 0.04934447258710861, 0.04784965515136719, 0.013765772804617882, 0.06403610110282898, 0.018437998369336128, -0.004653866868466139, 0.017492491751909256, -0.10154706239700317, 0.030901899561285973, -0.056882862001657486, -9.263546414476393e-33, 0.032071489840745926, -0.0631452277302742, 0.08019828796386719, 0.046986088156700134, -0.041777417063713074, 0.019012004137039185, 0.0045827883295714855, -0.0044275070540606976, -0.048470113426446915, -0.04983275383710861, -0.01822718232870102, -0.056053366512060165, -0.01828877255320549, -0.08254925161600113, 0.09881918132305145, 0.007796112447977066, -0.06468319147825241, 0.0760025680065155, -0.060067180544137955, 0.013709616847336292, 0.03374325856566429, 0.08879794180393219, 0.024127846583724022, -0.00841542985290289, 0.02462981455028057, 0.021595649421215057, 0.006472499575465918, -0.04018458351492882, 0.056987080723047256, 0.027295934036374092, 0.043269768357276917, -0.035261642187833786, -0.0553387887775898, 0.0012502741301432252, 0.022229338064789772, 0.03574218973517418, -0.030987700447440147, -0.05300057306885719, 0.03411632776260376, 0.03243895620107651, 0.01464895997196436, -0.03079126589000225, 0.012365912087261677, -0.023611238226294518, -0.019613271579146385, 0.020287813618779182, -0.05531042069196701, 0.007631160318851471, 0.04959418624639511, -0.0028270920738577843, -0.03970704972743988, -0.0012680422514677048, -0.10289273411035538, -0.016428861767053604, 0.051626332104206085, -0.024168673902750015, 0.01413167454302311, 0.004920460749417543, 0.06266112625598907, 0.08981794118881226, -0.03504524007439613, 0.10948540270328522, -0.030017409473657608, 0.05491165444254875, -0.028453201055526733, 0.03521046042442322, 0.03874867781996727, -0.028874889016151428, -0.029068557545542717, 0.009280937723815441, 0.03872179985046387, -0.013307967223227024, 0.025576138868927956, -0.059037335216999054, -0.02324475534260273, -0.018154621124267578, 0.05024878308176994, -0.03812074288725853, -0.0005742490175180137, 0.0959213599562645, -0.08256079256534576, 0.0028051540721207857, 0.07616480439901352, 0.0471491701900959, -0.032695621252059937, 0.047330521047115326, -0.03588801249861717, -0.001659886329434812, 0.010642332956194878, -0.006945728324353695, -0.09805209934711456, -0.05619858577847481, 0.027882959693670273, -0.05604751035571098, -0.10592898726463318, 4.5949629689841814e-33, -0.02613542228937149, -0.006271560676395893, 0.0457206629216671, 0.03334523364901543, 0.03378026932477951, -0.026292745023965836, 0.04472527652978897, 0.17767280340194702, -0.03315893933176994, 0.004677652381360531, 0.09173073619604111, -0.016673699021339417, 0.1146891638636589, 0.04301730915904045, 0.06296279281377792, 0.06419584900140762, 0.08006896078586578, -0.0630093514919281, -0.04692116379737854, -0.0552690327167511, -0.051715556532144547, 0.07841462641954422, 7.64712312957272e-05, 0.07047218084335327, -0.029466819018125534, 0.0026049974840134382, 0.010392696596682072, -0.02673671767115593, -0.12546704709529877, -0.05680812895298004, -0.004450392443686724, -0.011376120150089264, -0.11601763963699341, 0.055631838738918304, -0.030161255970597267, 0.05621485784649849, -0.009667915292084217, 0.014201724901795387, 0.011887979693710804, 0.11088187992572784, 0.054768797010183334, -0.011493510566651821, 0.015432603657245636, 0.15411017835140228, 0.02968861348927021, 0.01719019003212452, -0.04063909873366356, 0.06752496212720871, 0.016008097678422928, -0.02203967794775963, -0.0654606819152832, 0.018488537520170212, -0.09390315413475037, 0.0002816113119479269, 0.036626268178224564, -0.0027083519380539656, -0.07971291989088058, 0.04305737838149071, -0.005723634269088507, -0.017149874940514565, 0.03781665116548538, 0.036178670823574066, -0.03408880531787872, 0.04199240356683731, -0.011985711753368378, 0.0028120034839957952, -0.03526388108730316, 0.06770520657300949, -0.08036894351243973, -0.042860716581344604, 0.08275144547224045, -0.0190457571297884, -0.06447432190179825, -0.0013508021365851164, -0.11597707122564316, 0.09675540775060654, 0.025637278333306313, 0.03058444894850254, 0.04587564244866371, -0.0248523261398077, -0.09393557906150818, -0.0026686012279242277, -0.019358353689312935, 0.014417261816561222, -0.029305467382073402, 0.004780230578035116, 0.06614120304584503, -0.11459170281887054, 0.030194466933608055, 0.044906068593263626, -0.043565358966588974, -0.02789449878036976, 0.029291005805134773, -0.08918238431215286, -0.021739071235060692, -4.051126722970366e-08, 0.050828926265239716, 0.09618842601776123, -0.04176934435963631, -0.04011300578713417, 0.05205736681818962, 0.03308804705739021, 0.06941086053848267, 0.03233039379119873, 0.019226500764489174, 0.028578905388712883, 0.08651749789714813, -0.026446888223290443, -0.006831962149590254, -0.010019184090197086, -0.025466298684477806, -0.07011319696903229, -0.03966142237186432, 0.09581855684518814, -0.010219096206128597, -0.010008341632783413, -0.006578824482858181, 0.047231514006853104, 0.06408018618822098, -0.09780948609113693, 0.01514657773077488, -0.02015356533229351, 0.004760416690260172, 0.01761423423886299, 0.010450637899339199, -0.014002174139022827, 0.0002521428687032312, -0.008097398094832897, -0.06720700114965439, -0.13718897104263306, -0.01689615473151207, 0.014587732031941414, -0.028353476896882057, -0.02413005568087101, -0.034959759563207626, -0.11913619935512543, -0.00029614943196065724, 0.05826498195528984, 0.01299167051911354, 0.0897693783044815, 0.07465356588363647, 0.00587701378390193, -0.03150487691164017, -0.05592454597353935, 0.03574699908494949, -0.1141921803355217, -0.105095773935318, 0.04447728767991066, 0.0626031756401062, -0.01961544156074524, -0.007970171049237251, -0.0004538485372904688, -0.01703469268977642, -0.05490802973508835, -0.050680164247751236, -0.035791438072919846, 0.09396317601203918, -0.14849130809307098, 0.08407998085021973, 0.06755759567022324], "stable": [-0.009878501296043396, 0.0660642459988594, -0.05964508652687073, 0.07210763543844223, -0.03596775606274605, 0.02193950116634369, 0.010246319696307182, 0.01852239854633808, -0.04952670633792877, 0.020290177315473557, 0.023801632225513458, -0.05430629104375839, -0.007544808555394411, 0.012260775081813335, -0.06389766931533813, 0.03241700306534767, -0.053984079509973526, -0.07338772714138031, 0.05813191086053848, -0.06823553144931793, -0.005922270938754082, -0.011817307211458683, 0.03510754182934761, -0.008638807572424412, 0.06499254703521729, -0.01711813174188137, -0.024374794214963913, 0.007834694348275661, -0.015754472464323044, -0.03877580538392067, -0.041879672557115555, -0.052929844707250595, 0.030507778748869896, 0.03479808196425438, 0.0003642712254077196, -0.0018468601629137993, 0.09256058931350708, 0.01781170256435871, 0.03229816257953644, 0.0028201970271766186, -0.04097898676991463, -0.05748302862048149, 0.058425866067409515, 0.035652417689561844, 0.017063038423657417, -0.00208694557659328, 1.945298572536558e-05, 0.020170750096440315, 0.0525299571454525, 0.06233098730444908, 0.01996224746108055, -0.007750102784484625, -0.05245819315314293, -0.031384505331516266, 0.07049878686666489, 0.01562911458313465, -0.025653736665844917, 0.005772517994046211, 0.033708617091178894, 0.0173160582780838, -0.038299381732940674, -0.023969877511262894, -0.08515048772096634, 0.03904426097869873, 0.0337010882794857, -0.01881220191717148, -0.022112177684903145, -0.0830058604478836, -0.01255874801427126, -0.04907488450407982, 0.058949124068021774, -0.021205447614192963, 0.025199897587299347, -0.06132317706942558, -0.004188921302556992, -0.014213490299880505, 0.045258406549692154, 0.0980139970779419, -0.03531423583626747, -0.030920423567295074, -0.03895387426018715, -0.06878765672445297, -0.01498069055378437, 0.041184403002262115, 0.024011759087443352, 0.024984082207083702, -0.02330908738076687, 0.04412960261106491, -0.03279155492782593, -0.009370105341076851, -0.04633424058556557, 0.00010355184349464253, -0.007409134414047003, 0.05980905145406723, -0.10712308436632156, -0.0016119017964228988, -0.02891424112021923, 0.05164550617337227, -0.01999475061893463, 0.08253007382154465, 0.013118532486259937, 0.0412047877907753, 0.06558967381715775, 0.04939449205994606, 0.012494037859141827, 0.03203149884939194, 0.021725697442889214, 0.04889465123414993, -0.005762160290032625, -0.006175832357257605, -0.05037158355116844, -0.03631116822361946, 0.022419270128011703, 0.05585220083594322, 0.05395171791315079, -0.019489850848913193, 0.023703837767243385, 0.05224357917904854, 0.036405641585588455, -0.04141676053404808, 0.054783858358860016, 0.011990341357886791, -0.026010261848568916, 0.002385602332651615, -0.16738146543502808, 0.053104545921087265, -0.04563811421394348, -1.1385195433591571e-32, 0.03814498707652092, -0.0689789205789566, 0.04041396453976631, 0.08191869407892227, 0.014672595076262951, 0.005293590947985649, -0.006078558508306742, 0.019997579976916313, -0.040207281708717346, -0.0526622049510479, -0.009530546143651009, -0.1254967600107193, -0.05126592144370079, -0.12994910776615143, 0.06820761412382126, 0.03678153455257416, -0.08823713660240173, 0.02777552604675293, -0.03936967998743057, -0.020527733489871025, 0.011278019286692142, 0.0680217295885086, -0.0006382729043252766, -0.05814585089683533, -0.023165149614214897, 0.015882395207881927, 0.017337651923298836, -0.050064388662576675, 0.03945594280958176, 0.03115176595747471, 0.026074353605508804, -0.03448919579386711, -0.030477076768875122, 0.009733759798109531, 0.06363176554441452, 0.04730990156531334, -0.038697969168424606, -0.09644670784473419, 0.005985030438750982, 0.06498679518699646, 0.07189147919416428, -0.013101630844175816, 0.007560222875326872, -0.009547903202474117, -0.010277465917170048, -0.022786013782024384, -0.028719522058963776, 0.023598508909344673, 0.11158104240894318, -0.007700982037931681, -0.04821554571390152, -0.005352020729333162, -0.07801919430494308, -0.02650800719857216, 0.04496317356824875, 0.011938092298805714, 0.026169680058956146, -0.0013595058117061853, 0.05485756695270538, 0.06567436456680298, 0.0012906172778457403, 0.13648147881031036, 0.023734988644719124, 0.08998269587755203, -0.0095582390204072, 0.026485173031687737, 0.02947866916656494, -0.03661210834980011, -0.04421375319361687, -0.017298301681876183, 0.02297912910580635, -0.0008622432942502201, 0.043219152837991714, -0.07702449709177017, -0.048244915902614594, -0.0011753002181649208, -0.007110916543751955, -0.053615424782037735, -0.02565167471766472, 0.03301990404725075, -0.0672312006354332, 0.02777446247637272, 0.033486258238554, 0.055901262909173965, -0.06408447027206421, -0.010863428935408592, -0.015546686016023159, -0.001255270792171359, -0.008552140556275845, -0.014277186244726181, -0.11752666532993317, -0.03857119381427765, 0.036908265203237534, -0.10304670035839081, -0.1406712532043457, 5.338419432183107e-33, -0.035239845514297485, -0.024217309430241585, 0.05473233014345169, 0.022820761427283287, 0.028393885120749474, -0.04210122302174568, 0.035068243741989136, 0.15622800588607788, 0.0221007838845253, 0.03761698678135872, 0.07671965658664703, -0.02328198216855526, 0.10679496824741364, 0.025215674191713333, 0.073450468480587, 0.05074753239750862, 0.06479737162590027, -0.003941272385418415, -0.04551118239760399, -0.049933552742004395, -0.07364815473556519, 0.02552993968129158, -0.0714607834815979, 0.015462617389857769, -0.012316695414483547, 0.054023656994104385, 0.06723211705684662, -0.042197808623313904, -0.13952405750751495, -0.04436447471380234, -0.002935767639428377, 0.024541763588786125, -0.10958870500326157, 0.03318606689572334, -0.03618679940700531, 0.09100988507270813, 0.0005205267225392163, -0.0015193012077361345, 0.02488815411925316, 0.06314726918935776, 0.01987849734723568, -0.008496267721056938, 0.004493177868425846, 0.2061968892812729, 0.02684602700173855, 0.020799856632947922, -0.0240657776594162, 0.10361810028553009, 0.031669653952121735, 0.011405418626964092, -0.08602043241262436, 0.03416803479194641, -0.006508660968393087, -0.018249591812491417, 0.01674700528383255, -0.013997111469507217, -0.05165862664580345, 0.04553484916687012, 0.028897026553750038, -0.022380631417036057, 0.01289576105773449, -0.008146153762936592, -0.014556516893208027, 0.018003376200795174, -0.004316835198551416, 0.01513249333947897, -0.04365140199661255, 0.12434182316064835, -0.0656566470861435, -0.03079024702310562, 0.12046895921230316, 0.04114476218819618, -0.034990835934877396, 0.007664693985134363, -0.09582898765802383, 0.05962927266955376, 0.029368393123149872, 0.02580225095152855, 0.007645400706678629, -0.027397513389587402, -0.1363711804151535, 0.022010598331689835, -0.03838039189577103, 0.038143340498209, -0.02666621282696724, -0.015414954163134098, 0.030951106920838356, -0.06612234562635422, 0.007744719740003347, -0.0015180594054982066, -0.04261331632733345, -0.010056674480438232, 0.027605708688497543, -0.09087405353784561, -0.024939894676208496, -4.877212234077888e-08, 0.02553410269320011, 0.1265815794467926, -0.04538703337311745, -0.025004295632243156, 0.014260372146964073, 0.025149479508399963, 0.07109005749225616, 0.03147389367222786, -0.009967900812625885, 0.044358670711517334, -0.0068642375990748405, 0.014319314621388912, 0.02237335965037346, -0.004660268314182758, -0.003350135637447238, -0.05754431337118149, -0.027631275355815887, 0.04403224214911461, -0.024214953184127808, 0.03008132055401802, 0.047187186777591705, -0.018362415954470634, 0.04596714302897453, -0.08293260633945465, -0.00023648078786209226, -0.01700695976614952, 0.02500813640654087, 0.036679912358522415, -0.0171666257083416, -0.012540695257484913, -0.010196696035563946, 0.013239305466413498, -0.07839847356081009, -0.10489100217819214, -0.029697928577661514, 0.004295755177736282, 0.0014388053677976131, -0.013160772621631622, -0.0455457828938961, -0.15142354369163513, 0.012144980952143669, 0.14141105115413666, 0.0033308325801044703, 0.09230928868055344, 0.09918193519115448, 0.02047782950103283, -0.0018556228606030345, -0.01753951981663704, 0.013949376530945301, -0.06404218077659607, -0.09226547181606293, 0.06893295794725418, 0.06536483019590378, -0.020992908626794815, -0.019392572343349457, -0.04842235520482063, -0.004990136716514826, -0.04778695106506348, -0.07618878036737442, 0.020379342138767242, 0.08492814004421234, -0.06385879963636398, 0.06381313502788544, 0.023798275738954544]}}

webfetch/extract/__init__.py

@@ -0,0 +1,35 @@
+"""
+Extract stage public interface.
+
+Takes the top chunks from the ranking stage, concatenates them into a
+token-budgeted context, calls an LLM, and returns structured values for the
+requested extraction keys.
+
+Pick an adapter explicitly:
+
+    from webfetch.extract import ClaudeExtractor
+    extractor = ClaudeExtractor()
+    values = extractor.extract(
+        chunks=top_chunks,
+        keys={
+            "accuracy": "measurement accuracy including units and conditions",
+            "range": "full measurement range with units",
+            "resolution": "smallest measurable increment with units",
+        },
+    )
+"""
+
+from webfetch.extract.base import AbstractExtractor, build_context
+from webfetch.extract.claude import ClaudeExtractor
+from webfetch.extract.gemini import GeminiExtractor
+from webfetch.extract.gpt import GPTExtractor
+from webfetch.extract.groq import GroqExtractor
+
+__all__ = [
+    "AbstractExtractor",
+    "ClaudeExtractor",
+    "GPTExtractor",
+    "GeminiExtractor",
+    "GroqExtractor",
+    "build_context",
+]