leansearch 0.2.0__tar.gz → 0.2.2__tar.gz

{leansearch-0.2.0 → leansearch-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: leansearch
-Version: 0.2.0
+Version: 0.2.2
 Summary: Embedded full-text and vector search in one library (LeanText + LeanVec).
 Author-email: Carlo Moro <cnmoro@gmail.com>
 License: MIT
@@ -13,8 +13,8 @@ Classifier: Programming Language :: Python :: 3
 Classifier: Topic :: Database
 Classifier: Topic :: Text Processing :: Indexing
 Requires-Python: >=3.8
-Requires-Dist: leantext<2.3,>=2.2.38
-Requires-Dist: leanvec-scale<1.1,>=1.0.44
+Requires-Dist: leantext<2.3,>=2.2.41
+Requires-Dist: leanvec-scale<1.1,>=1.0.45
 Requires-Dist: numpy
 Requires-Dist: orjson
 Provides-Extra: grpc

{leansearch-0.2.0 → leansearch-0.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "leansearch"
-version = "0.2.0"
+version = "0.2.2"
 description = "Embedded full-text and vector search in one library (LeanText + LeanVec)."
 requires-python = ">=3.8"
 license = { text = "MIT" }
@@ -14,8 +14,8 @@ dependencies = [
     # so it pins a compatible release range checked by tests/test_internals.py.
     # The servers (leansearch.server) reuse both engines' server guardrails, which ship
     # in their wheels from these releases on.
-    "leanvec-scale>=1.0.44,<1.1",
-    "leantext>=2.2.38,<2.3",
+    "leanvec-scale>=1.0.45,<1.1",
+    "leantext>=2.2.41,<2.3",
     "numpy",
     "orjson",
 ]

{leansearch-0.2.0 → leansearch-0.2.2}/src/leansearch/__init__.py

@@ -15,5 +15,5 @@ from ._collection import Collection, Snapshot
 from ._db import LeanSearch
 from ._table import Table
 
-__version__ = "0.2.0"
+__version__ = "0.2.2"
 __all__ = ["LeanSearch", "Collection", "Snapshot", "Table", "__version__"]

{leansearch-0.2.0 → leansearch-0.2.2}/src/leansearch/_collection.py

@@ -12,12 +12,15 @@ import time
 from typing import Any, Dict, Iterable, List, Optional, Sequence
 
 from . import _vectors
-from ._fusion import rrf_fuse
+from ._fusion import check_fusion, relative_fuse, rrf_fuse
 
 # Keyword arguments each engine's search accepts, used to split a hybrid call's options.
 _TEXT_SEARCH_OPTIONS = frozenset({
     "fuzzy_dist", "title_boost", "body_boost", "return_content", "return_highlight",
     "highlight_buffer", "match_any", "force_fresh", "max_result_bytes"})
+# Text options only LeanText's ranked `search` takes: a hybrid query passing one uses that
+# ranking on the text side instead of BM25.
+_RANKED_ONLY_OPTIONS = frozenset({"fuzzy_dist", "title_boost", "body_boost"})
 _VECTOR_SEARCH_OPTIONS = frozenset({
     "include_metadata", "ensure_target_k", "nprobe", "rerank", "max_result_bytes"})
 
@@ -456,7 +459,15 @@ class Collection:
                 if text_rows:
                     t._check_field_sizes([docs[i].get("text") or "" for i in text_rows],
                                          prepared)
-            if len(explicit) <= _PROBE_IDS_DIRECTLY:
+            probe = getattr(t, "_exists_for_write", None)
+            if probe is not None:
+                # leantext >= 2.2.41: answered by the writable handle under the write
+                # lock, not by a read-only handle that must replay this process's writes.
+                f = {"id": {"$in": explicit}}
+                if probe(f, collection=self.name):
+                    sizes_ok()
+                    t.delete_by_filter(f, collection=self.name)
+            elif len(explicit) <= _PROBE_IDS_DIRECTLY:
                 old = [i for i in explicit if t.get_content(i, collection=self.name) is not None]
                 if old:
                     sizes_ok()
@@ -660,8 +671,9 @@ class Collection:
 
     # ---- search --------------------------------------------------------------------
     def search(self, query: Optional[str] = None, *, vector=None, sparse=None, k: int = 10,
-               filters: Optional[Dict[str, Any]] = None, rrf_k: float = 60,
+               filters: Optional[Dict[str, Any]] = None, rrf_k: Optional[float] = None,
                page: Optional[int] = None, page_size: Optional[int] = None,
+               fusion: Optional[str] = None, alpha: Optional[float] = None,
                **options) -> List[Dict[str, Any]]:
         """Search by text, by vector, or by both.
 
@@ -672,10 +684,20 @@ class Collection:
           (`LeanVecDB.search`, hybrid dense+sparse when `sparse` is given); options
           `nprobe`, `rerank`, `ensure_target_k`, `include_metadata`. Score: squared L2
           distance, lower is better (dense+sparse: RRF, higher is better).
-        - `search("words", vector=v)`: both rankings fused by reciprocal rank (the same
-          rule LeanVec uses for dense+sparse): score `sum(1 / (rrf_k + rank))`, higher is
-          better; each hit also carries `text_score`/`text_rank` and
-          `vector_score`/`vector_rank` where it was found.
+        - `search("words", vector=v)`: hybrid. The text side is BM25 (LeanText's
+          `search_bm25`, any query word matches unless `match_any=False`); with
+          `fuzzy_dist`, `title_boost` or `body_boost` it is LeanText's `search` ranking
+          instead. Each side ranks max(10 x depth, 100) candidates under the same filter,
+          and the two rankings are fused:
+          `fusion="relative"` (default): each side's scores are min-max normalized to
+          [0, 1] against its top 100 hits (deeper hits clamp to 0) and a document scores `(1 - alpha) * text + alpha * vector`, `alpha`
+          default 0.75 (the vector side's weight).
+          `fusion="rrf"` (also chosen by passing `rrf_k`): reciprocal rank, the rule
+          LeanVec uses for dense+sparse, score `sum(1 / (rrf_k + rank))`, `rrf_k`
+          default 60.
+          Higher is better; each hit also carries `text_score`/`text_rank` and
+          `vector_score`/`vector_rank` where it was found. A text-only or vector-only
+          search ignores `fusion` and `alpha` (as a text-only search ignores `rrf_k`).
 
         `page`/`page_size` (0-based) return one page of the ranking instead of the top k.
         `k` defaults to 10 here for every kind of query (`LeanVecDB.search` alone defaults
@@ -696,20 +718,24 @@ class Collection:
         if has_vector_query and not has_text_query:
             return self._vec.search(vector, k=k, filters=self._vector_filter(filters),
                                     collection=self.name,
-                                    query_sparse_embedding=sparse, rrf_k=rrf_k,
+                                    query_sparse_embedding=sparse,
+                                    rrf_k=60 if rrf_k is None else rrf_k,
                                     page_size=page_size,
                                     page_index=page if page_size is not None else None,
                                     **options)
         return self._hybrid(query, vector, sparse, k, filters, rrf_k, page, page_size,
-                            options)
+                            options, fusion, alpha)
 
-    def _hybrid(self, query, vector, sparse, k, filters, rrf_k, page, page_size, options):
+    def _hybrid(self, query, vector, sparse, k, filters, rrf_k, page, page_size, options,
+                fusion=None, alpha=None):
         # Both engines are needed: without one, the "fused" ranking silently was the
         # other engine's alone (a vector-only search on a text-only store raises).
         self._text, self._vec
         unknown = set(options) - _TEXT_SEARCH_OPTIONS - _VECTOR_SEARCH_OPTIONS
         if unknown:
             raise TypeError(f"unexpected search options {sorted(unknown)}")
+        fusion, alpha = check_fusion(fusion, alpha, rrf_k)
+        rrf_k = 60 if rrf_k is None else rrf_k
         text_opts = {o: v for o, v in options.items() if o in _TEXT_SEARCH_OPTIONS}
         vec_opts = {o: v for o, v in options.items() if o in _VECTOR_SEARCH_OPTIONS}
         if page_size is not None:
@@ -724,13 +750,28 @@ class Collection:
         fetch = max(total * 10, 100)
         if filters:
             filters = self._text._sanitize_filter(filters)   # once, for both engines
-        text_hits = (self._text.search(query, k=fetch, filters=filters, collection=self.name,
-                                       **text_opts) if self._has_text() else [])
+        if self._has_text():
+            if _RANKED_ONLY_OPTIONS.intersection(text_opts):
+                text_hits = self._text.search(query, k=fetch, filters=filters,
+                                              collection=self.name, **text_opts)
+            else:
+                text_opts.setdefault("match_any", True)
+                text_hits = self._text.search_bm25(query=query, k=fetch, filters=filters,
+                                                   collection=self.name, **text_opts)
+        else:
+            text_hits = []
         vec_hits = (self._vec.search(vector, k=fetch, filters=self._vector_filter(filters),
                                      collection=self.name,
                                      query_sparse_embedding=sparse, rrf_k=rrf_k, **vec_opts)
                     if self._has_vectors() else [])
-        fused = rrf_fuse(text_hits, vec_hits, rrf_k)[start:total]
+        if fusion == "rrf":
+            fused = rrf_fuse(text_hits, vec_hits, rrf_k)[start:total]
+        else:
+            # LeanVec fuses dense + sparse (RRF, higher is better) only for a non-empty
+            # sparse query; otherwise its scores are distances (lower is better).
+            sparse_fused = sparse is not None and bool(_vectors._to_sparse(sparse))
+            fused = relative_fuse(text_hits, vec_hits, alpha,
+                                  vector_higher_is_better=sparse_fused)[start:total]
         if vec_opts.get("include_metadata") is False:
             for h in fused:
                 h.pop("metadata", None)

leansearch-0.2.2/src/leansearch/_fusion.py

@@ -0,0 +1,148 @@
+"""Reciprocal-rank fusion of a text ranking and a vector ranking."""
+from typing import Any, Dict, List
+
+
+def rrf_fuse(text_hits: List[Dict[str, Any]], vector_hits: List[Dict[str, Any]],
+             rrf_k: float) -> List[Dict[str, Any]]:
+    """Fuse two rankings the way LeanVec fuses dense and sparse: each list adds
+    `1 / (rrf_k + rank)` (rank from 0) to a document's score; the result is sorted by
+    score, descending, and ties keep first-seen order (text hits, then vector-only hits).
+
+    A fused hit is the text hit's fields (title, content, highlight, metadata) or, for a
+    document only the vector search found, the vector hit's, plus `score` (the fused
+    score) and `text_score`/`text_rank`/`vector_score`/`vector_rank` where it was found.
+    The hit dicts are annotated in place: callers pass the lists an engine search just
+    returned, which nothing else holds.
+    """
+    import math
+    rrf_k = float(rrf_k)
+    if not math.isfinite(rrf_k) or rrf_k <= 0:
+        raise ValueError(f"rrf_k must be a finite positive number, got {rrf_k!r}")
+    fused: Dict[str, Dict[str, Any]] = {}
+    for rank, h in enumerate(text_hits):
+        rid = h["id"]
+        if rid not in fused:          # the engines return each id once; keep the first
+            h["text_score"] = h.get("score")
+            h["text_rank"] = rank
+            h["score"] = 1.0 / (rrf_k + rank)
+            fused[rid] = h
+    seen = set()
+    for rank, h in enumerate(vector_hits):
+        rid = h["id"]
+        if rid in seen:
+            continue
+        seen.add(rid)
+        hit = fused.get(rid)
+        if hit is None:
+            h["vector_score"] = h.get("score")
+            h["vector_rank"] = rank
+            h["score"] = 1.0 / (rrf_k + rank)
+            fused[rid] = h
+        else:
+            hit["vector_score"] = h.get("score")
+            hit["vector_rank"] = rank
+            hit["score"] += 1.0 / (rrf_k + rank)
+    out = list(fused.values())
+    out.sort(key=_neg_score)          # stable: ties keep first-seen order
+    return out
+
+
+def _neg_score(h):
+    return -h["score"]
+
+
+# Hybrid fusion defaults (Collection.search with a text query AND a vector).
+DEFAULT_FUSION = "relative"
+DEFAULT_ALPHA = 0.75          # weight of the vector side in relative-score fusion
+FUSIONS = ("relative", "rrf")
+NORM_WINDOW = 100             # relative fusion normalizes against each side's top 100
+
+
+def check_fusion(fusion, alpha, rrf_k):
+    """Resolve the fusion method and its weight.
+
+    `fusion=None` means the default, relative-score fusion, unless `rrf_k` is given: an
+    explicit `rrf_k` keeps the earlier reciprocal-rank fusion. `alpha` (0..1) is the
+    vector side's weight and only applies to relative-score fusion."""
+    import math
+    if fusion is None:
+        fusion = "rrf" if rrf_k is not None else DEFAULT_FUSION
+    if fusion not in FUSIONS:
+        raise ValueError(f"fusion must be one of {FUSIONS}, got {fusion!r}")
+    if alpha is None:
+        alpha = DEFAULT_ALPHA
+    else:
+        if fusion != "relative":
+            raise ValueError("alpha applies to fusion='relative' only")
+        try:
+            alpha = float(alpha)
+        except (TypeError, ValueError):
+            raise ValueError(f"alpha must be a number in [0, 1], got {alpha!r}") from None
+        if not math.isfinite(alpha) or not 0.0 <= alpha <= 1.0:
+            raise ValueError(f"alpha must be a number in [0, 1], got {alpha!r}")
+    return fusion, alpha
+
+
+def _minmax(hits, higher_is_better, window):
+    """Each hit's score mapped to [0, 1] against the best and the worst score among the
+    list's first `window` hits (best = 1; deeper hits clamp to 0). The window is fixed
+    (not the read depth), so a hit's normalized score does not depend on how deep each
+    side was read or which page was asked for. A window whose scores are all equal maps
+    to 1, except all-zero BM25 scores (no text signal), which map to 0."""
+    if not hits:
+        return []
+    xs = [float(h.get("score")) for h in hits]
+    ref = xs[:window]
+    lo, hi = min(ref), max(ref)
+    span = hi - lo
+    if not span > 0:
+        if higher_is_better and not hi > 0:
+            # e.g. an empty text query browsing a filter: every BM25 score is 0, no signal
+            return [0.0] * len(xs)
+        return [1.0 if i < window else 0.0 for i in range(len(xs))]
+    if higher_is_better:
+        return [min(1.0, max(0.0, (x - lo) / span)) for x in xs]
+    return [min(1.0, max(0.0, (hi - x) / span)) for x in xs]
+
+
+def relative_fuse(text_hits: List[Dict[str, Any]], vector_hits: List[Dict[str, Any]],
+                  alpha: float, vector_higher_is_better: bool,
+                  window: int = NORM_WINDOW) -> List[Dict[str, Any]]:
+    """Relative-score (convex) fusion: each list's scores are min-max normalized to
+    [0, 1] over its first `window` hits (best hit = 1, see `_minmax`), and a document
+    scores `(1 - alpha) * text + alpha * vector` (a side that did not return the
+    document adds 0). Sorted by that score, descending;
+    ties keep first-seen order (text hits, then vector-only hits).
+
+    Text scores are BM25 (higher is better); vector scores are distances (lower is
+    better) unless the vector side itself fused dense and sparse by RRF. Same hit fields
+    as `rrf_fuse`: `score` (fused) plus `text_score`/`text_rank` and
+    `vector_score`/`vector_rank` where the document was found."""
+    wt, wv = 1.0 - alpha, alpha
+    fused: Dict[str, Dict[str, Any]] = {}
+    for rank, (h, x) in enumerate(zip(text_hits, _minmax(text_hits, True, window))):
+        rid = h["id"]
+        if rid not in fused:
+            h["text_score"] = h.get("score")
+            h["text_rank"] = rank
+            h["score"] = wt * x
+            fused[rid] = h
+    seen = set()
+    for rank, (h, x) in enumerate(zip(vector_hits, _minmax(vector_hits, vector_higher_is_better, window))):
+        rid = h["id"]
+        if rid in seen:
+            continue
+        seen.add(rid)
+        hit = fused.get(rid)
+        if hit is None:
+            h["vector_score"] = h.get("score")
+            h["vector_rank"] = rank
+            h["score"] = wv * x
+            fused[rid] = h
+        else:
+            hit["vector_score"] = h.get("score")
+            hit["vector_rank"] = rank
+            hit["score"] += wv * x
+    out = list(fused.values())
+    out.sort(key=_neg_score)
+    return out

{leansearch-0.2.0 → leansearch-0.2.2}/src/leansearch/server/_service.py

@@ -23,6 +23,7 @@ from leanvec.server import serve_guardrails as _vg
 from leantext_server import serve_guardrails as _tg
 
 from .. import LeanSearch
+from .._fusion import check_fusion
 
 GuardrailError = _vg.GuardrailError
 InvalidRequestError = _vg.InvalidRequestError
@@ -319,8 +320,10 @@ class Service:
             tg.check_fanout(many, per)
 
     def search(self, collection, *, query=None, vector=None, sparse=None, k=None,
-               filters=None, page=None, page_size=None, rrf_k=None, **opts):
-        """`Collection.search`: text, vector (+ sparse), or both fused by RRF."""
+               filters=None, page=None, page_size=None, rrf_k=None, fusion=None,
+               alpha=None, **opts):
+        """`Collection.search`: text, vector (+ sparse), or both fused (relative score by
+        default, or RRF)."""
         unknown = set(opts) - set(_TEXT_OPTIONS) - set(_VECTOR_OPTIONS)
         if unknown:
             raise InvalidRequestError(f"unknown search options {sorted(unknown)}")
@@ -359,11 +362,23 @@ class Service:
             if rrf_k is not None:
                 # finite and > 0: rrf_k <= 0 divides by zero inside the fusion
                 rrf_k = vg.check_rrf_k(rrf_k)
+            if has_text:
+                try:
+                    check_fusion(fusion or None, alpha, rrf_k)
+                except ValueError as e:
+                    raise InvalidRequestError(str(e)) from None
+            else:
+                fusion = alpha = None    # a vector-only search ignores them, as the library does
         else:
             rrf_k = None                 # a text-only search ignores it, as the library does
+            fusion = alpha = None
         call["max_result_bytes"] = vg.result_byte_budget()
         if rrf_k is not None:
             call["rrf_k"] = rrf_k
+        if fusion:
+            call["fusion"] = fusion
+        if alpha is not None:
+            call["alpha"] = alpha
         return self.db[_name(collection)].search(
             query, vector=vector, sparse=sparse, k=k, filters=filters, page=page,
             page_size=page_size, **call)

{leansearch-0.2.0 → leansearch-0.2.2}/src/leansearch/server/client.py

@@ -293,6 +293,7 @@ class LeanSearchClient:
     def search(self, query: Optional[str] = None, *, vector=None, sparse=None, k: int = 10,
                filters: Optional[Dict[str, Any]] = None, page: Optional[int] = None,
                page_size: Optional[int] = None, rrf_k: Optional[float] = None,
+               fusion: Optional[str] = None, alpha: Optional[float] = None,
                fuzzy_dist: Optional[int] = None, title_boost: Optional[float] = None,
                body_boost: Optional[float] = None, return_content: Optional[bool] = None,
                return_highlight: Optional[bool] = None,
@@ -301,12 +302,13 @@ class LeanSearchClient:
                ensure_target_k: Optional[bool] = None,
                include_metadata: Optional[bool] = None, force_fresh: Optional[bool] = None,
                collection: str = "default", timeout=None) -> List[Dict[str, Any]]:
-        """Search by text, by vector (+ sparse), or both fused by RRF (Collection.search)."""
+        """Search by text, by vector (+ sparse), or both fused (Collection.search)."""
         req = pb.SearchRequest(collection=collection, vector=_vec(vector),
                                filters_json=_filters(filters),
                                nprobe=nprobe or 0, rerank=rerank or 0)
         _set(req, k=k, page_size=page_size, force_fresh=force_fresh,
-             query=query, page=page, rrf_k=rrf_k, fuzzy_dist=fuzzy_dist,
+             query=query, page=page, rrf_k=rrf_k, fusion=fusion, alpha=alpha,
+             fuzzy_dist=fuzzy_dist,
              title_boost=title_boost, body_boost=body_boost, return_content=return_content,
              return_highlight=return_highlight, highlight_buffer=highlight_buffer,
              match_any=match_any, ensure_target_k=ensure_target_k,

{leansearch-0.2.0 → leansearch-0.2.2}/src/leansearch/server/leansearch.proto

@@ -154,7 +154,7 @@ message SearchRequest {
   string filters_json = 6;
   optional int32 page = 7;              // 0-based page of the ranking (with page_size)
   optional int32 page_size = 8;         // unset -> top k, no paging
-  optional double rrf_k = 9;            // fusion constant (default 60)
+  optional double rrf_k = 9;            // RRF constant (default 60); set -> fusion rrf
   // text options
   optional int32 fuzzy_dist = 10;
   optional double title_boost = 11;
@@ -169,6 +169,9 @@ message SearchRequest {
   optional bool ensure_target_k = 19;
   optional bool include_metadata = 20;  // default true
   optional bool force_fresh = 21;       // text: reload from disk before searching
+  // hybrid fusion (text query + vector): "relative" (default; rrf when rrf_k is set) or "rrf"
+  optional string fusion = 22;
+  optional double alpha = 23;           // relative fusion: vector weight in [0, 1], default 0.75
 }
 
 message QueryRequest {

leansearch-0.2.2/src/leansearch/server/leansearch_pb2.py

@@ -0,0 +1,108 @@
+# -*- coding: utf-8 -*-
+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# NO CHECKED-IN PROTOBUF GENCODE
+# source: leansearch.proto
+# Protobuf Python Version: 7.35.0
+"""Generated protocol buffer code."""
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import descriptor_pool as _descriptor_pool
+from google.protobuf import runtime_version as _runtime_version
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf.internal import builder as _builder
+_runtime_version.ValidateProtobufRuntimeVersion(
+    _runtime_version.Domain.PUBLIC,
+    7,
+    35,
+    0,
+    '',
+    'leansearch.proto'
+)
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+
+
+
+DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\x10leansearch.proto\x12\nleansearch\"\x07\n\x05\x45mpty\"/\n\x0cSparseVector\x12\x0f\n\x07indices\x18\x01 \x03(\r\x12\x0e\n\x06values\x18\x02 \x03(\x02\"\x0f\n\rHealthRequest\"1\n\x0bHealthReply\x12\r\n\x05ready\x18\x01 \x01(\x08\x12\x13\n\x0blimits_json\x18\x02 \x01(\t\"\'\n\x11\x43ollectionRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\"\x1a\n\tListReply\x12\r\n\x05names\x18\x01 \x03(\t\"\x1b\n\nCountReply\x12\r\n\x05\x63ount\x18\x01 \x01(\x03\"\x19\n\tJsonReply\x12\x0c\n\x04json\x18\x01 \x01(\t\"\x1d\n\x0bStatusReply\x12\x0e\n\x06status\x18\x01 \x01(\t\"\x15\n\x07IdReply\x12\n\n\x02id\x18\x01 \x01(\t\"\x17\n\x08IdsReply\x12\x0b\n\x03ids\x18\x01 \x03(\t\"\x1d\n\x0b\x45xistsReply\x12\x0e\n\x06\x65xists\x18\x01 \x01(\x08\" \n\x0bSearchReply\x12\x11\n\thits_json\x18\x01 \x01(\t\"\xa1\x01\n\x08\x44ocument\x12\n\n\x02id\x18\x01 \x01(\t\x12\x11\n\x04text\x18\x02 \x01(\tH\x00\x88\x01\x01\x12\x12\n\x05title\x18\x03 \x01(\tH\x01\x88\x01\x01\x12\x0e\n\x06vector\x18\x04 \x03(\x02\x12(\n\x06sparse\x18\x05 \x01(\x0b\x32\x18.leansearch.SparseVector\x12\x15\n\rmetadata_json\x18\x06 \x01(\tB\x07\n\x05_textB\x08\n\x06_title\"\\\n\nAddRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12&\n\x08\x64ocument\x18\x02 \x01(\x0b\x32\x14.leansearch.Document\x12\x12\n\nskip_index\x18\x03 \x03(\t\"a\n\x0e\x41\x64\x64ManyRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\'\n\tdocuments\x18\x02 \x03(\x0b\x32\x14.leansearch.Document\x12\x12\n\nskip_index\x18\x03 \x03(\t\"E\n\nGetRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x0b\n\x03ids\x18\x02 \x03(\t\x12\x16\n\x0einclude_vector\x18\x03 \x01(\x08\"0\n\x0e\x43ontentRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\n\n\x02id\x18\x02 \x01(\t\"0\n\x0c\x43ontentReply\x12\x14\n\x07\x63ontent\x18\x01 \x01(\tH\x00\x88\x01\x01\x42\n\n\x08_content\"0\n\rDeleteRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x0b\n\x03ids\x18\x02 \x03(\t\"9\n\rFilterRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x02 \x01(\t\"V\n\x0cItemsRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x02 \x01(\t\x12\x12\n\x05limit\x18\x03 \x01(\x05H\x00\x88\x01\x01\x42\x08\n\x06_limit\"\xad\x06\n\rSearchRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x12\n\x05query\x18\x02 \x01(\tH\x00\x88\x01\x01\x12\x0e\n\x06vector\x18\x03 \x03(\x02\x12(\n\x06sparse\x18\x04 \x01(\x0b\x32\x18.leansearch.SparseVector\x12\x0e\n\x01k\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x06 \x01(\t\x12\x11\n\x04page\x18\x07 \x01(\x05H\x02\x88\x01\x01\x12\x16\n\tpage_size\x18\x08 \x01(\x05H\x03\x88\x01\x01\x12\x12\n\x05rrf_k\x18\t \x01(\x01H\x04\x88\x01\x01\x12\x17\n\nfuzzy_dist\x18\n \x01(\x05H\x05\x88\x01\x01\x12\x18\n\x0btitle_boost\x18\x0b \x01(\x01H\x06\x88\x01\x01\x12\x17\n\nbody_boost\x18\x0c \x01(\x01H\x07\x88\x01\x01\x12\x1b\n\x0ereturn_content\x18\r \x01(\x08H\x08\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x0e \x01(\x08H\t\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\x0f \x01(\x05H\n\x88\x01\x01\x12\x16\n\tmatch_any\x18\x10 \x01(\x08H\x0b\x88\x01\x01\x12\x0e\n\x06nprobe\x18\x11 \x01(\x05\x12\x0e\n\x06rerank\x18\x12 \x01(\x05\x12\x1c\n\x0f\x65nsure_target_k\x18\x13 \x01(\x08H\x0c\x88\x01\x01\x12\x1d\n\x10include_metadata\x18\x14 \x01(\x08H\r\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\x15 \x01(\x08H\x0e\x88\x01\x01\x12\x13\n\x06\x66usion\x18\x16 \x01(\tH\x0f\x88\x01\x01\x12\x12\n\x05\x61lpha\x18\x17 \x01(\x01H\x10\x88\x01\x01\x42\x08\n\x06_queryB\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\x08\n\x06_rrf_kB\r\n\x0b_fuzzy_distB\x0e\n\x0c_title_boostB\r\n\x0b_body_boostB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0c\n\n_match_anyB\x12\n\x10_ensure_target_kB\x13\n\x11_include_metadataB\x0e\n\x0c_force_freshB\t\n\x07_fusionB\x08\n\x06_alpha\"\xcd\x03\n\x0cQueryRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05query\x18\x02 \x01(\t\x12\x0e\n\x01k\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x11\n\x04page\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x16\n\tpage_size\x18\x06 \x01(\x05H\x02\x88\x01\x01\x12\x18\n\x0btitle_boost\x18\x07 \x01(\x01H\x03\x88\x01\x01\x12\x17\n\nbody_boost\x18\x08 \x01(\x01H\x04\x88\x01\x01\x12\x18\n\x10\x64\x65\x66\x61ult_operator\x18\t \x01(\t\x12\x1b\n\x0ereturn_content\x18\n \x01(\x08H\x05\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x0b \x01(\x08H\x06\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\x0c \x01(\x05H\x07\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\r \x01(\x08H\x08\x88\x01\x01\x42\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\x0e\n\x0c_title_boostB\r\n\x0b_body_boostB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0e\n\x0c_force_fresh\"\x86\x03\n\x0b\x42m25Request\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05query\x18\x02 \x01(\t\x12\x0e\n\x01k\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x11\n\x04page\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x16\n\tpage_size\x18\x06 \x01(\x05H\x02\x88\x01\x01\x12\x1b\n\x0ereturn_content\x18\x07 \x01(\x08H\x03\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x08 \x01(\x08H\x04\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\t \x01(\x05H\x05\x88\x01\x01\x12\x16\n\tmatch_any\x18\n \x01(\x08H\x06\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\x0b \x01(\x08H\x07\x88\x01\x01\x42\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0c\n\n_match_anyB\x0e\n\x0c_force_fresh\"\x88\x04\n\x11SearchManyRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x0f\n\x07queries\x18\x02 \x03(\t\x12\x0e\n\x01k\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x11\n\x04page\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x16\n\tpage_size\x18\x06 \x01(\x05H\x02\x88\x01\x01\x12\x17\n\nfuzzy_dist\x18\x07 \x01(\x05H\x03\x88\x01\x01\x12\x18\n\x0btitle_boost\x18\x08 \x01(\x01H\x04\x88\x01\x01\x12\x17\n\nbody_boost\x18\t \x01(\x01H\x05\x88\x01\x01\x12\x1b\n\x0ereturn_content\x18\n \x01(\x08H\x06\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x0b \x01(\x08H\x07\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\x0c \x01(\x05H\x08\x88\x01\x01\x12\x16\n\tmatch_any\x18\r \x01(\x08H\t\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\x0e \x01(\x08H\n\x88\x01\x01\x42\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\r\n\x0b_fuzzy_distB\x0e\n\x0c_title_boostB\r\n\x0b_body_boostB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0c\n\n_match_anyB\x0e\n\x0c_force_fresh\"M\n\x12ValueCountsRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05\x66ield\x18\x02 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x03 \x01(\t\"\x9f\x01\n\x11TopByGroupRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x10\n\x08group_by\x18\x02 \x01(\t\x12\x0f\n\x07sort_by\x18\x03 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x15\n\routput_fields\x18\x05 \x03(\t\x12\x17\n\ndescending\x18\x06 \x01(\x08H\x00\x88\x01\x01\x42\r\n\x0b_descending\"[\n\x0cPca3dRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x12\n\nbatch_size\x18\x02 \x01(\x05\x12\x14\n\x0c\x66ilters_json\x18\x03 \x01(\t\x12\r\n\x05limit\x18\x04 \x01(\x05\"P\n\nPca3dPoint\x12\n\n\x02id\x18\x01 \x01(\t\x12\t\n\x01x\x18\x02 \x01(\x01\x12\t\n\x01y\x18\x03 \x01(\x01\x12\t\n\x01z\x18\x04 \x01(\x01\x12\x15\n\rmetadata_json\x18\x05 \x01(\t\"2\n\rVerifyRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05\x66lush\x18\x02 \x01(\x08\"!\n\x0bLvqlRequest\x12\x12\n\nquery_json\x18\x01 \x01(\t\"L\n\x10TableRowsRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x11\n\trows_json\x18\x02 \x01(\t\x12\x16\n\x0evector_columns\x18\x03 \x03(\t\"6\n\x11TableQueryRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x12\n\nquery_json\x18\x02 \x01(\t\"o\n\x12TableSearchRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x0e\n\x06\x63olumn\x18\x02 \x01(\t\x12\x0e\n\x06vector\x18\x03 \x03(\x02\x12\x0e\n\x01k\x18\x04 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x05 \x01(\tB\x04\n\x02_k\"2\n\x11TableCountRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x0e\n\x06\x63olumn\x18\x02 \x01(\t2\xb9\x10\n\nLeanSearch\x12<\n\x06Health\x12\x19.leansearch.HealthRequest\x1a\x17.leansearch.HealthReply\x12;\n\x0fListCollections\x12\x11.leansearch.Empty\x1a\x15.leansearch.ListReply\x12>\n\x05\x43ount\x12\x1d.leansearch.CollectionRequest\x1a\x16.leansearch.CountReply\x12=\n\x05Stats\x12\x1d.leansearch.CollectionRequest\x1a\x15.leansearch.JsonReply\x12@\n\x08\x44\x65scribe\x12\x1d.leansearch.CollectionRequest\x1a\x15.leansearch.JsonReply\x12\x32\n\x03\x41\x64\x64\x12\x16.leansearch.AddRequest\x1a\x13.leansearch.IdReply\x12;\n\x07\x41\x64\x64Many\x12\x1a.leansearch.AddManyRequest\x1a\x14.leansearch.IdsReply\x12\x34\n\x03Get\x12\x16.leansearch.GetRequest\x1a\x15.leansearch.JsonReply\x12?\n\x07\x43ontent\x12\x1a.leansearch.ContentRequest\x1a\x18.leansearch.ContentReply\x12;\n\x06\x44\x65lete\x12\x19.leansearch.DeleteRequest\x1a\x16.leansearch.CountReply\x12@\n\x0b\x44\x65leteWhere\x12\x19.leansearch.FilterRequest\x1a\x16.leansearch.CountReply\x12>\n\x05\x43lear\x12\x1d.leansearch.CollectionRequest\x1a\x16.leansearch.CountReply\x12<\n\x06\x45xists\x12\x19.leansearch.FilterRequest\x1a\x17.leansearch.ExistsReply\x12\x38\n\x05Items\x12\x18.leansearch.ItemsRequest\x1a\x15.leansearch.JsonReply\x12<\n\x06Search\x12\x19.leansearch.SearchRequest\x1a\x17.leansearch.SearchReply\x12:\n\x05Query\x12\x18.leansearch.QueryRequest\x1a\x17.leansearch.SearchReply\x12\x38\n\x04\x42m25\x12\x17.leansearch.Bm25Request\x1a\x17.leansearch.SearchReply\x12\x42\n\nSearchMany\x12\x1d.leansearch.SearchManyRequest\x1a\x15.leansearch.JsonReply\x12\x44\n\x0bValueCounts\x12\x1e.leansearch.ValueCountsRequest\x1a\x15.leansearch.JsonReply\x12\x42\n\nTopByGroup\x12\x1d.leansearch.TopByGroupRequest\x1a\x15.leansearch.JsonReply\x12;\n\x05Pca3d\x12\x18.leansearch.Pca3dRequest\x1a\x16.leansearch.Pca3dPoint0\x01\x12\x41\n\x07Persist\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12\x46\n\x0c\x46lushBarrier\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12\x41\n\x07Refresh\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12@\n\x06Vacuum\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12:\n\x06Verify\x12\x19.leansearch.VerifyRequest\x1a\x15.leansearch.JsonReply\x12\x36\n\x04Lvql\x12\x17.leansearch.LvqlRequest\x1a\x15.leansearch.JsonReply\x12\x36\n\nListTables\x12\x11.leansearch.Empty\x1a\x15.leansearch.JsonReply\x12@\n\x08TableAdd\x12\x1c.leansearch.TableRowsRequest\x1a\x16.leansearch.CountReply\x12\x44\n\x0cTableReplace\x12\x1c.leansearch.TableRowsRequest\x1a\x16.leansearch.CountReply\x12\x42\n\nTableQuery\x12\x1d.leansearch.TableQueryRequest\x1a\x15.leansearch.JsonReply\x12\x44\n\x0bTableSearch\x12\x1e.leansearch.TableSearchRequest\x1a\x15.leansearch.JsonReply\x12\x43\n\nTableCount\x12\x1d.leansearch.TableCountRequest\x1a\x16.leansearch.CountReplyb\x06proto3')
+
+_globals = globals()
+_builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
+_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, 'leansearch_pb2', _globals)
+if not _descriptor._USE_C_DESCRIPTORS:
+  DESCRIPTOR._loaded_options = None
+  _globals['_EMPTY']._serialized_start=32
+  _globals['_EMPTY']._serialized_end=39
+  _globals['_SPARSEVECTOR']._serialized_start=41
+  _globals['_SPARSEVECTOR']._serialized_end=88
+  _globals['_HEALTHREQUEST']._serialized_start=90
+  _globals['_HEALTHREQUEST']._serialized_end=105
+  _globals['_HEALTHREPLY']._serialized_start=107
+  _globals['_HEALTHREPLY']._serialized_end=156
+  _globals['_COLLECTIONREQUEST']._serialized_start=158
+  _globals['_COLLECTIONREQUEST']._serialized_end=197
+  _globals['_LISTREPLY']._serialized_start=199
+  _globals['_LISTREPLY']._serialized_end=225
+  _globals['_COUNTREPLY']._serialized_start=227
+  _globals['_COUNTREPLY']._serialized_end=254
+  _globals['_JSONREPLY']._serialized_start=256
+  _globals['_JSONREPLY']._serialized_end=281
+  _globals['_STATUSREPLY']._serialized_start=283
+  _globals['_STATUSREPLY']._serialized_end=312
+  _globals['_IDREPLY']._serialized_start=314
+  _globals['_IDREPLY']._serialized_end=335
+  _globals['_IDSREPLY']._serialized_start=337
+  _globals['_IDSREPLY']._serialized_end=360
+  _globals['_EXISTSREPLY']._serialized_start=362
+  _globals['_EXISTSREPLY']._serialized_end=391
+  _globals['_SEARCHREPLY']._serialized_start=393
+  _globals['_SEARCHREPLY']._serialized_end=425
+  _globals['_DOCUMENT']._serialized_start=428
+  _globals['_DOCUMENT']._serialized_end=589
+  _globals['_ADDREQUEST']._serialized_start=591
+  _globals['_ADDREQUEST']._serialized_end=683
+  _globals['_ADDMANYREQUEST']._serialized_start=685
+  _globals['_ADDMANYREQUEST']._serialized_end=782
+  _globals['_GETREQUEST']._serialized_start=784
+  _globals['_GETREQUEST']._serialized_end=853
+  _globals['_CONTENTREQUEST']._serialized_start=855
+  _globals['_CONTENTREQUEST']._serialized_end=903
+  _globals['_CONTENTREPLY']._serialized_start=905
+  _globals['_CONTENTREPLY']._serialized_end=953
+  _globals['_DELETEREQUEST']._serialized_start=955
+  _globals['_DELETEREQUEST']._serialized_end=1003
+  _globals['_FILTERREQUEST']._serialized_start=1005
+  _globals['_FILTERREQUEST']._serialized_end=1062
+  _globals['_ITEMSREQUEST']._serialized_start=1064
+  _globals['_ITEMSREQUEST']._serialized_end=1150
+  _globals['_SEARCHREQUEST']._serialized_start=1153
+  _globals['_SEARCHREQUEST']._serialized_end=1966
+  _globals['_QUERYREQUEST']._serialized_start=1969
+  _globals['_QUERYREQUEST']._serialized_end=2430
+  _globals['_BM25REQUEST']._serialized_start=2433
+  _globals['_BM25REQUEST']._serialized_end=2823
+  _globals['_SEARCHMANYREQUEST']._serialized_start=2826
+  _globals['_SEARCHMANYREQUEST']._serialized_end=3346
+  _globals['_VALUECOUNTSREQUEST']._serialized_start=3348
+  _globals['_VALUECOUNTSREQUEST']._serialized_end=3425
+  _globals['_TOPBYGROUPREQUEST']._serialized_start=3428
+  _globals['_TOPBYGROUPREQUEST']._serialized_end=3587
+  _globals['_PCA3DREQUEST']._serialized_start=3589
+  _globals['_PCA3DREQUEST']._serialized_end=3680
+  _globals['_PCA3DPOINT']._serialized_start=3682
+  _globals['_PCA3DPOINT']._serialized_end=3762
+  _globals['_VERIFYREQUEST']._serialized_start=3764
+  _globals['_VERIFYREQUEST']._serialized_end=3814
+  _globals['_LVQLREQUEST']._serialized_start=3816
+  _globals['_LVQLREQUEST']._serialized_end=3849
+  _globals['_TABLEROWSREQUEST']._serialized_start=3851
+  _globals['_TABLEROWSREQUEST']._serialized_end=3927
+  _globals['_TABLEQUERYREQUEST']._serialized_start=3929
+  _globals['_TABLEQUERYREQUEST']._serialized_end=3983
+  _globals['_TABLESEARCHREQUEST']._serialized_start=3985
+  _globals['_TABLESEARCHREQUEST']._serialized_end=4096
+  _globals['_TABLECOUNTREQUEST']._serialized_start=4098
+  _globals['_TABLECOUNTREQUEST']._serialized_end=4148
+  _globals['_LEANSEARCH']._serialized_start=4151
+  _globals['_LEANSEARCH']._serialized_end=6256
+# @@protoc_insertion_point(module_scope)

{leansearch-0.2.0 → leansearch-0.2.2}/src/leansearch/server/serve.py

@@ -385,6 +385,8 @@ class SearchReq(_TextOpts):
     page: Optional[int] = None
     page_size: Optional[int] = None
     rrf_k: Optional[float] = None
+    fusion: Optional[str] = None
+    alpha: Optional[float] = None
     fuzzy_dist: Optional[int] = None
     title_boost: Optional[float] = None
     body_boost: Optional[float] = None
@@ -560,7 +562,7 @@ def search(req: SearchReq):
     return {"hits": _call(lambda: svc().search(
         req.collection, query=req.query, vector=req.vector, sparse=req.sparse, k=req.k,
         filters=req.filters, page=req.page, page_size=req.page_size, rrf_k=req.rrf_k,
-        **opts))}
+        fusion=req.fusion, alpha=req.alpha, **opts))}
 
 
 @app.post("/query")

{leansearch-0.2.0 → leansearch-0.2.2}/src/leansearch/server/serve_grpc.py

@@ -297,7 +297,8 @@ class LeanSearchServicer(pb_grpc.LeanSearchServicer):
                 r.collection, query=r.query if r.HasField("query") else None,
                 vector=list(r.vector) if len(r.vector) else None, sparse=_sparse(r.sparse),
                 k=_has(r, "k"), filters=_json(r.filters_json), page=_has(r, "page"),
-                page_size=_has(r, "page_size"), rrf_k=_has(r, "rrf_k"), **opts)
+                page_size=_has(r, "page_size"), rrf_k=_has(r, "rrf_k"),
+                fusion=_has(r, "fusion"), alpha=_has(r, "alpha"), **opts)
             return pb.SearchReply(hits_json=S.dumps(hits))
         return self._run(context, pb.SearchReply, run,
                          S.json_weight(request.filters_json)

leansearch-0.2.0/src/leansearch/_fusion.py

@@ -1,51 +0,0 @@
-"""Reciprocal-rank fusion of a text ranking and a vector ranking."""
-from typing import Any, Dict, List
-
-
-def rrf_fuse(text_hits: List[Dict[str, Any]], vector_hits: List[Dict[str, Any]],
-             rrf_k: float) -> List[Dict[str, Any]]:
-    """Fuse two rankings the way LeanVec fuses dense and sparse: each list adds
-    `1 / (rrf_k + rank)` (rank from 0) to a document's score; the result is sorted by
-    score, descending, and ties keep first-seen order (text hits, then vector-only hits).
-
-    A fused hit is the text hit's fields (title, content, highlight, metadata) or, for a
-    document only the vector search found, the vector hit's, plus `score` (the fused
-    score) and `text_score`/`text_rank`/`vector_score`/`vector_rank` where it was found.
-    The hit dicts are annotated in place: callers pass the lists an engine search just
-    returned, which nothing else holds.
-    """
-    import math
-    rrf_k = float(rrf_k)
-    if not math.isfinite(rrf_k) or rrf_k <= 0:
-        raise ValueError(f"rrf_k must be a finite positive number, got {rrf_k!r}")
-    fused: Dict[str, Dict[str, Any]] = {}
-    for rank, h in enumerate(text_hits):
-        rid = h["id"]
-        if rid not in fused:          # the engines return each id once; keep the first
-            h["text_score"] = h.get("score")
-            h["text_rank"] = rank
-            h["score"] = 1.0 / (rrf_k + rank)
-            fused[rid] = h
-    seen = set()
-    for rank, h in enumerate(vector_hits):
-        rid = h["id"]
-        if rid in seen:
-            continue
-        seen.add(rid)
-        hit = fused.get(rid)
-        if hit is None:
-            h["vector_score"] = h.get("score")
-            h["vector_rank"] = rank
-            h["score"] = 1.0 / (rrf_k + rank)
-            fused[rid] = h
-        else:
-            hit["vector_score"] = h.get("score")
-            hit["vector_rank"] = rank
-            hit["score"] += 1.0 / (rrf_k + rank)
-    out = list(fused.values())
-    out.sort(key=_neg_score)          # stable: ties keep first-seen order
-    return out
-
-
-def _neg_score(h):
-    return -h["score"]

leansearch-0.2.0/src/leansearch/server/leansearch_pb2.py

@@ -1,108 +0,0 @@
-# -*- coding: utf-8 -*-
-# Generated by the protocol buffer compiler.  DO NOT EDIT!
-# NO CHECKED-IN PROTOBUF GENCODE
-# source: leansearch.proto
-# Protobuf Python Version: 7.35.0
-"""Generated protocol buffer code."""
-from google.protobuf import descriptor as _descriptor
-from google.protobuf import descriptor_pool as _descriptor_pool
-from google.protobuf import runtime_version as _runtime_version
-from google.protobuf import symbol_database as _symbol_database
-from google.protobuf.internal import builder as _builder
-_runtime_version.ValidateProtobufRuntimeVersion(
-    _runtime_version.Domain.PUBLIC,
-    7,
-    35,
-    0,
-    '',
-    'leansearch.proto'
-)
-# @@protoc_insertion_point(imports)
-
-_sym_db = _symbol_database.Default()
-
-
-
-
-DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\x10leansearch.proto\x12\nleansearch\"\x07\n\x05\x45mpty\"/\n\x0cSparseVector\x12\x0f\n\x07indices\x18\x01 \x03(\r\x12\x0e\n\x06values\x18\x02 \x03(\x02\"\x0f\n\rHealthRequest\"1\n\x0bHealthReply\x12\r\n\x05ready\x18\x01 \x01(\x08\x12\x13\n\x0blimits_json\x18\x02 \x01(\t\"\'\n\x11\x43ollectionRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\"\x1a\n\tListReply\x12\r\n\x05names\x18\x01 \x03(\t\"\x1b\n\nCountReply\x12\r\n\x05\x63ount\x18\x01 \x01(\x03\"\x19\n\tJsonReply\x12\x0c\n\x04json\x18\x01 \x01(\t\"\x1d\n\x0bStatusReply\x12\x0e\n\x06status\x18\x01 \x01(\t\"\x15\n\x07IdReply\x12\n\n\x02id\x18\x01 \x01(\t\"\x17\n\x08IdsReply\x12\x0b\n\x03ids\x18\x01 \x03(\t\"\x1d\n\x0b\x45xistsReply\x12\x0e\n\x06\x65xists\x18\x01 \x01(\x08\" \n\x0bSearchReply\x12\x11\n\thits_json\x18\x01 \x01(\t\"\xa1\x01\n\x08\x44ocument\x12\n\n\x02id\x18\x01 \x01(\t\x12\x11\n\x04text\x18\x02 \x01(\tH\x00\x88\x01\x01\x12\x12\n\x05title\x18\x03 \x01(\tH\x01\x88\x01\x01\x12\x0e\n\x06vector\x18\x04 \x03(\x02\x12(\n\x06sparse\x18\x05 \x01(\x0b\x32\x18.leansearch.SparseVector\x12\x15\n\rmetadata_json\x18\x06 \x01(\tB\x07\n\x05_textB\x08\n\x06_title\"\\\n\nAddRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12&\n\x08\x64ocument\x18\x02 \x01(\x0b\x32\x14.leansearch.Document\x12\x12\n\nskip_index\x18\x03 \x03(\t\"a\n\x0e\x41\x64\x64ManyRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\'\n\tdocuments\x18\x02 \x03(\x0b\x32\x14.leansearch.Document\x12\x12\n\nskip_index\x18\x03 \x03(\t\"E\n\nGetRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x0b\n\x03ids\x18\x02 \x03(\t\x12\x16\n\x0einclude_vector\x18\x03 \x01(\x08\"0\n\x0e\x43ontentRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\n\n\x02id\x18\x02 \x01(\t\"0\n\x0c\x43ontentReply\x12\x14\n\x07\x63ontent\x18\x01 \x01(\tH\x00\x88\x01\x01\x42\n\n\x08_content\"0\n\rDeleteRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x0b\n\x03ids\x18\x02 \x03(\t\"9\n\rFilterRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x02 \x01(\t\"V\n\x0cItemsRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x02 \x01(\t\x12\x12\n\x05limit\x18\x03 \x01(\x05H\x00\x88\x01\x01\x42\x08\n\x06_limit\"\xef\x05\n\rSearchRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x12\n\x05query\x18\x02 \x01(\tH\x00\x88\x01\x01\x12\x0e\n\x06vector\x18\x03 \x03(\x02\x12(\n\x06sparse\x18\x04 \x01(\x0b\x32\x18.leansearch.SparseVector\x12\x0e\n\x01k\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x06 \x01(\t\x12\x11\n\x04page\x18\x07 \x01(\x05H\x02\x88\x01\x01\x12\x16\n\tpage_size\x18\x08 \x01(\x05H\x03\x88\x01\x01\x12\x12\n\x05rrf_k\x18\t \x01(\x01H\x04\x88\x01\x01\x12\x17\n\nfuzzy_dist\x18\n \x01(\x05H\x05\x88\x01\x01\x12\x18\n\x0btitle_boost\x18\x0b \x01(\x01H\x06\x88\x01\x01\x12\x17\n\nbody_boost\x18\x0c \x01(\x01H\x07\x88\x01\x01\x12\x1b\n\x0ereturn_content\x18\r \x01(\x08H\x08\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x0e \x01(\x08H\t\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\x0f \x01(\x05H\n\x88\x01\x01\x12\x16\n\tmatch_any\x18\x10 \x01(\x08H\x0b\x88\x01\x01\x12\x0e\n\x06nprobe\x18\x11 \x01(\x05\x12\x0e\n\x06rerank\x18\x12 \x01(\x05\x12\x1c\n\x0f\x65nsure_target_k\x18\x13 \x01(\x08H\x0c\x88\x01\x01\x12\x1d\n\x10include_metadata\x18\x14 \x01(\x08H\r\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\x15 \x01(\x08H\x0e\x88\x01\x01\x42\x08\n\x06_queryB\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\x08\n\x06_rrf_kB\r\n\x0b_fuzzy_distB\x0e\n\x0c_title_boostB\r\n\x0b_body_boostB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0c\n\n_match_anyB\x12\n\x10_ensure_target_kB\x13\n\x11_include_metadataB\x0e\n\x0c_force_fresh\"\xcd\x03\n\x0cQueryRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05query\x18\x02 \x01(\t\x12\x0e\n\x01k\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x11\n\x04page\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x16\n\tpage_size\x18\x06 \x01(\x05H\x02\x88\x01\x01\x12\x18\n\x0btitle_boost\x18\x07 \x01(\x01H\x03\x88\x01\x01\x12\x17\n\nbody_boost\x18\x08 \x01(\x01H\x04\x88\x01\x01\x12\x18\n\x10\x64\x65\x66\x61ult_operator\x18\t \x01(\t\x12\x1b\n\x0ereturn_content\x18\n \x01(\x08H\x05\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x0b \x01(\x08H\x06\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\x0c \x01(\x05H\x07\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\r \x01(\x08H\x08\x88\x01\x01\x42\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\x0e\n\x0c_title_boostB\r\n\x0b_body_boostB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0e\n\x0c_force_fresh\"\x86\x03\n\x0b\x42m25Request\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05query\x18\x02 \x01(\t\x12\x0e\n\x01k\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x11\n\x04page\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x16\n\tpage_size\x18\x06 \x01(\x05H\x02\x88\x01\x01\x12\x1b\n\x0ereturn_content\x18\x07 \x01(\x08H\x03\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x08 \x01(\x08H\x04\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\t \x01(\x05H\x05\x88\x01\x01\x12\x16\n\tmatch_any\x18\n \x01(\x08H\x06\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\x0b \x01(\x08H\x07\x88\x01\x01\x42\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0c\n\n_match_anyB\x0e\n\x0c_force_fresh\"\x88\x04\n\x11SearchManyRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x0f\n\x07queries\x18\x02 \x03(\t\x12\x0e\n\x01k\x18\x03 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x11\n\x04page\x18\x05 \x01(\x05H\x01\x88\x01\x01\x12\x16\n\tpage_size\x18\x06 \x01(\x05H\x02\x88\x01\x01\x12\x17\n\nfuzzy_dist\x18\x07 \x01(\x05H\x03\x88\x01\x01\x12\x18\n\x0btitle_boost\x18\x08 \x01(\x01H\x04\x88\x01\x01\x12\x17\n\nbody_boost\x18\t \x01(\x01H\x05\x88\x01\x01\x12\x1b\n\x0ereturn_content\x18\n \x01(\x08H\x06\x88\x01\x01\x12\x1d\n\x10return_highlight\x18\x0b \x01(\x08H\x07\x88\x01\x01\x12\x1d\n\x10highlight_buffer\x18\x0c \x01(\x05H\x08\x88\x01\x01\x12\x16\n\tmatch_any\x18\r \x01(\x08H\t\x88\x01\x01\x12\x18\n\x0b\x66orce_fresh\x18\x0e \x01(\x08H\n\x88\x01\x01\x42\x04\n\x02_kB\x07\n\x05_pageB\x0c\n\n_page_sizeB\r\n\x0b_fuzzy_distB\x0e\n\x0c_title_boostB\r\n\x0b_body_boostB\x11\n\x0f_return_contentB\x13\n\x11_return_highlightB\x13\n\x11_highlight_bufferB\x0c\n\n_match_anyB\x0e\n\x0c_force_fresh\"M\n\x12ValueCountsRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05\x66ield\x18\x02 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x03 \x01(\t\"\x9f\x01\n\x11TopByGroupRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x10\n\x08group_by\x18\x02 \x01(\t\x12\x0f\n\x07sort_by\x18\x03 \x01(\t\x12\x14\n\x0c\x66ilters_json\x18\x04 \x01(\t\x12\x15\n\routput_fields\x18\x05 \x03(\t\x12\x17\n\ndescending\x18\x06 \x01(\x08H\x00\x88\x01\x01\x42\r\n\x0b_descending\"[\n\x0cPca3dRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\x12\n\nbatch_size\x18\x02 \x01(\x05\x12\x14\n\x0c\x66ilters_json\x18\x03 \x01(\t\x12\r\n\x05limit\x18\x04 \x01(\x05\"P\n\nPca3dPoint\x12\n\n\x02id\x18\x01 \x01(\t\x12\t\n\x01x\x18\x02 \x01(\x01\x12\t\n\x01y\x18\x03 \x01(\x01\x12\t\n\x01z\x18\x04 \x01(\x01\x12\x15\n\rmetadata_json\x18\x05 \x01(\t\"2\n\rVerifyRequest\x12\x12\n\ncollection\x18\x01 \x01(\t\x12\r\n\x05\x66lush\x18\x02 \x01(\x08\"!\n\x0bLvqlRequest\x12\x12\n\nquery_json\x18\x01 \x01(\t\"L\n\x10TableRowsRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x11\n\trows_json\x18\x02 \x01(\t\x12\x16\n\x0evector_columns\x18\x03 \x03(\t\"6\n\x11TableQueryRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x12\n\nquery_json\x18\x02 \x01(\t\"o\n\x12TableSearchRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x0e\n\x06\x63olumn\x18\x02 \x01(\t\x12\x0e\n\x06vector\x18\x03 \x03(\x02\x12\x0e\n\x01k\x18\x04 \x01(\x05H\x00\x88\x01\x01\x12\x14\n\x0c\x66ilters_json\x18\x05 \x01(\tB\x04\n\x02_k\"2\n\x11TableCountRequest\x12\r\n\x05table\x18\x01 \x01(\t\x12\x0e\n\x06\x63olumn\x18\x02 \x01(\t2\xb9\x10\n\nLeanSearch\x12<\n\x06Health\x12\x19.leansearch.HealthRequest\x1a\x17.leansearch.HealthReply\x12;\n\x0fListCollections\x12\x11.leansearch.Empty\x1a\x15.leansearch.ListReply\x12>\n\x05\x43ount\x12\x1d.leansearch.CollectionRequest\x1a\x16.leansearch.CountReply\x12=\n\x05Stats\x12\x1d.leansearch.CollectionRequest\x1a\x15.leansearch.JsonReply\x12@\n\x08\x44\x65scribe\x12\x1d.leansearch.CollectionRequest\x1a\x15.leansearch.JsonReply\x12\x32\n\x03\x41\x64\x64\x12\x16.leansearch.AddRequest\x1a\x13.leansearch.IdReply\x12;\n\x07\x41\x64\x64Many\x12\x1a.leansearch.AddManyRequest\x1a\x14.leansearch.IdsReply\x12\x34\n\x03Get\x12\x16.leansearch.GetRequest\x1a\x15.leansearch.JsonReply\x12?\n\x07\x43ontent\x12\x1a.leansearch.ContentRequest\x1a\x18.leansearch.ContentReply\x12;\n\x06\x44\x65lete\x12\x19.leansearch.DeleteRequest\x1a\x16.leansearch.CountReply\x12@\n\x0b\x44\x65leteWhere\x12\x19.leansearch.FilterRequest\x1a\x16.leansearch.CountReply\x12>\n\x05\x43lear\x12\x1d.leansearch.CollectionRequest\x1a\x16.leansearch.CountReply\x12<\n\x06\x45xists\x12\x19.leansearch.FilterRequest\x1a\x17.leansearch.ExistsReply\x12\x38\n\x05Items\x12\x18.leansearch.ItemsRequest\x1a\x15.leansearch.JsonReply\x12<\n\x06Search\x12\x19.leansearch.SearchRequest\x1a\x17.leansearch.SearchReply\x12:\n\x05Query\x12\x18.leansearch.QueryRequest\x1a\x17.leansearch.SearchReply\x12\x38\n\x04\x42m25\x12\x17.leansearch.Bm25Request\x1a\x17.leansearch.SearchReply\x12\x42\n\nSearchMany\x12\x1d.leansearch.SearchManyRequest\x1a\x15.leansearch.JsonReply\x12\x44\n\x0bValueCounts\x12\x1e.leansearch.ValueCountsRequest\x1a\x15.leansearch.JsonReply\x12\x42\n\nTopByGroup\x12\x1d.leansearch.TopByGroupRequest\x1a\x15.leansearch.JsonReply\x12;\n\x05Pca3d\x12\x18.leansearch.Pca3dRequest\x1a\x16.leansearch.Pca3dPoint0\x01\x12\x41\n\x07Persist\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12\x46\n\x0c\x46lushBarrier\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12\x41\n\x07Refresh\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12@\n\x06Vacuum\x12\x1d.leansearch.CollectionRequest\x1a\x17.leansearch.StatusReply\x12:\n\x06Verify\x12\x19.leansearch.VerifyRequest\x1a\x15.leansearch.JsonReply\x12\x36\n\x04Lvql\x12\x17.leansearch.LvqlRequest\x1a\x15.leansearch.JsonReply\x12\x36\n\nListTables\x12\x11.leansearch.Empty\x1a\x15.leansearch.JsonReply\x12@\n\x08TableAdd\x12\x1c.leansearch.TableRowsRequest\x1a\x16.leansearch.CountReply\x12\x44\n\x0cTableReplace\x12\x1c.leansearch.TableRowsRequest\x1a\x16.leansearch.CountReply\x12\x42\n\nTableQuery\x12\x1d.leansearch.TableQueryRequest\x1a\x15.leansearch.JsonReply\x12\x44\n\x0bTableSearch\x12\x1e.leansearch.TableSearchRequest\x1a\x15.leansearch.JsonReply\x12\x43\n\nTableCount\x12\x1d.leansearch.TableCountRequest\x1a\x16.leansearch.CountReplyb\x06proto3')
-
-_globals = globals()
-_builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
-_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, 'leansearch_pb2', _globals)
-if not _descriptor._USE_C_DESCRIPTORS:
-  DESCRIPTOR._loaded_options = None
-  _globals['_EMPTY']._serialized_start=32
-  _globals['_EMPTY']._serialized_end=39
-  _globals['_SPARSEVECTOR']._serialized_start=41
-  _globals['_SPARSEVECTOR']._serialized_end=88
-  _globals['_HEALTHREQUEST']._serialized_start=90
-  _globals['_HEALTHREQUEST']._serialized_end=105
-  _globals['_HEALTHREPLY']._serialized_start=107
-  _globals['_HEALTHREPLY']._serialized_end=156
-  _globals['_COLLECTIONREQUEST']._serialized_start=158
-  _globals['_COLLECTIONREQUEST']._serialized_end=197
-  _globals['_LISTREPLY']._serialized_start=199
-  _globals['_LISTREPLY']._serialized_end=225
-  _globals['_COUNTREPLY']._serialized_start=227
-  _globals['_COUNTREPLY']._serialized_end=254
-  _globals['_JSONREPLY']._serialized_start=256
-  _globals['_JSONREPLY']._serialized_end=281
-  _globals['_STATUSREPLY']._serialized_start=283
-  _globals['_STATUSREPLY']._serialized_end=312
-  _globals['_IDREPLY']._serialized_start=314
-  _globals['_IDREPLY']._serialized_end=335
-  _globals['_IDSREPLY']._serialized_start=337
-  _globals['_IDSREPLY']._serialized_end=360
-  _globals['_EXISTSREPLY']._serialized_start=362
-  _globals['_EXISTSREPLY']._serialized_end=391
-  _globals['_SEARCHREPLY']._serialized_start=393
-  _globals['_SEARCHREPLY']._serialized_end=425
-  _globals['_DOCUMENT']._serialized_start=428
-  _globals['_DOCUMENT']._serialized_end=589
-  _globals['_ADDREQUEST']._serialized_start=591
-  _globals['_ADDREQUEST']._serialized_end=683
-  _globals['_ADDMANYREQUEST']._serialized_start=685
-  _globals['_ADDMANYREQUEST']._serialized_end=782
-  _globals['_GETREQUEST']._serialized_start=784
-  _globals['_GETREQUEST']._serialized_end=853
-  _globals['_CONTENTREQUEST']._serialized_start=855
-  _globals['_CONTENTREQUEST']._serialized_end=903
-  _globals['_CONTENTREPLY']._serialized_start=905
-  _globals['_CONTENTREPLY']._serialized_end=953
-  _globals['_DELETEREQUEST']._serialized_start=955
-  _globals['_DELETEREQUEST']._serialized_end=1003
-  _globals['_FILTERREQUEST']._serialized_start=1005
-  _globals['_FILTERREQUEST']._serialized_end=1062
-  _globals['_ITEMSREQUEST']._serialized_start=1064
-  _globals['_ITEMSREQUEST']._serialized_end=1150
-  _globals['_SEARCHREQUEST']._serialized_start=1153
-  _globals['_SEARCHREQUEST']._serialized_end=1904
-  _globals['_QUERYREQUEST']._serialized_start=1907
-  _globals['_QUERYREQUEST']._serialized_end=2368
-  _globals['_BM25REQUEST']._serialized_start=2371
-  _globals['_BM25REQUEST']._serialized_end=2761
-  _globals['_SEARCHMANYREQUEST']._serialized_start=2764
-  _globals['_SEARCHMANYREQUEST']._serialized_end=3284
-  _globals['_VALUECOUNTSREQUEST']._serialized_start=3286
-  _globals['_VALUECOUNTSREQUEST']._serialized_end=3363
-  _globals['_TOPBYGROUPREQUEST']._serialized_start=3366
-  _globals['_TOPBYGROUPREQUEST']._serialized_end=3525
-  _globals['_PCA3DREQUEST']._serialized_start=3527
-  _globals['_PCA3DREQUEST']._serialized_end=3618
-  _globals['_PCA3DPOINT']._serialized_start=3620
-  _globals['_PCA3DPOINT']._serialized_end=3700
-  _globals['_VERIFYREQUEST']._serialized_start=3702
-  _globals['_VERIFYREQUEST']._serialized_end=3752
-  _globals['_LVQLREQUEST']._serialized_start=3754
-  _globals['_LVQLREQUEST']._serialized_end=3787
-  _globals['_TABLEROWSREQUEST']._serialized_start=3789
-  _globals['_TABLEROWSREQUEST']._serialized_end=3865
-  _globals['_TABLEQUERYREQUEST']._serialized_start=3867
-  _globals['_TABLEQUERYREQUEST']._serialized_end=3921
-  _globals['_TABLESEARCHREQUEST']._serialized_start=3923
-  _globals['_TABLESEARCHREQUEST']._serialized_end=4034
-  _globals['_TABLECOUNTREQUEST']._serialized_start=4036
-  _globals['_TABLECOUNTREQUEST']._serialized_end=4086
-  _globals['_LEANSEARCH']._serialized_start=4089
-  _globals['_LEANSEARCH']._serialized_end=6194
-# @@protoc_insertion_point(module_scope)