Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

Refactor: expose knn top_k and knn num_candidaties (#18737)

Wang Qi92c69db79e46
diff --git a/agent/tools/retrieval.py b/agent/tools/retrieval.py
index 81d5b286d7..190d8f5f81 100644
--- a/agent/tools/retrieval.py
+++ b/agent/tools/retrieval.py
@@ -189,7 +189,7 @@ class Retrieval(ToolBase, ABC):
self._param.top_n,
self._param.similarity_threshold,
1 - self._param.keywords_similarity_weight,
- top=self._param.top_k,
+ knn_top_k=self._param.top_k,
doc_ids=doc_ids,
aggs=True,
rerank_mdl=rerank_mdl,
diff --git a/api/apps/restful_apis/bot_api.py b/api/apps/restful_apis/bot_api.py
index 453e1f88c5..cb2da1d456 100644
--- a/api/apps/restful_apis/bot_api.py
+++ b/api/apps/restful_apis/bot_api.py
@@ -445,8 +445,8 @@ async def retrieval_test_embedded(tenant_id=None):
size,
similarity_threshold,
vector_similarity_weight,
- top,
- local_doc_ids,
+ doc_ids=local_doc_ids,
+ knn_top_k=top,
rerank_mdl=rerank_mdl,
highlight=req.get("highlight"),
rank_feature=labels,
diff --git a/api/apps/restful_apis/chunk_api.py b/api/apps/restful_apis/chunk_api.py
index 6ed59b24db..a18a6dc10f 100644
--- a/api/apps/restful_apis/chunk_api.py
+++ b/api/apps/restful_apis/chunk_api.py
@@ -419,8 +419,8 @@ async def retrieval_test(tenant_id):
size,
similarity_threshold,
vector_similarity_weight,
- top,
- doc_ids,
+ doc_ids=doc_ids,
+ knn_top_k=top,
rerank_mdl=rerank_mdl,
highlight=highlight,
rank_feature=label_question(question, kbs),
diff --git a/api/apps/restful_apis/dataset_api.py b/api/apps/restful_apis/dataset_api.py
index 341f01b683..4f956df457 100644
--- a/api/apps/restful_apis/dataset_api.py
+++ b/api/apps/restful_apis/dataset_api.py
@@ -513,9 +513,11 @@ async def search_datasets(tenant_id):
"""Search (retrieval test) across multiple datasets.
POST /api/v1/datasets/search
- JSON body: {"dataset_ids": list[str] (required), "question": str (required), "doc_ids": list[str], "top_k": int, "page": int, "size": int,
+ JSON body: {"dataset_ids": list[str] (required), "question": str (required), "doc_ids": list[str], "knn_top_k": int (default 1024), "knn_num_candidates": int (default 2048), "page": int, "size": int,
"similarity_threshold": float, "vector_similarity_weight": float, "use_kg": bool,
"cross_languages": list[str], "keyword": bool, "meta_data_filter": dict, "include_knowledge_compilation": bool (default true)}
+ The legacy "top_k" parameter is accepted as an alias for "knn_top_k".
+ "knn_num_candidates" currently applies only to Elasticsearch.
Success: {"code": 0, "data": {"chunks": [...], "total": int, "labels": [...]}}
Errors: ARGUMENT_ERROR (101) for invalid payload; DATA_ERROR (102) for access denied or internal errors.
"""
@@ -536,9 +538,11 @@ async def search(tenant_id, dataset_id):
"""Search (retrieval test) within a dataset.
POST /api/v1/datasets/<dataset_id>/search
- JSON body: {"question": str (required), "doc_ids": list[str], "top_k": int, "page": int, "size": int,
+ JSON body: {"question": str (required), "doc_ids": list[str], "knn_top_k": int (default 1024), "knn_num_candidates": int (default 2048), "page": int, "size": int,
"similarity_threshold": float, "vector_similarity_weight": float, "use_kg": bool,
"cross_languages": list[str], "keyword": bool, "meta_data_filter": dict, "include_knowledge_compilation": bool (default true)}
+ The legacy "top_k" parameter is accepted as an alias for "knn_top_k".
+ "knn_num_candidates" currently applies only to Elasticsearch.
Success: {"code": 0, "data": {"chunks": [...], "total": int, "labels": [...]}}
Errors: ARGUMENT_ERROR (101) for invalid payload; DATA_ERROR (102) for access denied or internal errors.
"""
diff --git a/api/apps/restful_apis/dify_retrieval_api.py b/api/apps/restful_apis/dify_retrieval_api.py
index c4eb214ecb..d271d9cdb6 100644
--- a/api/apps/restful_apis/dify_retrieval_api.py
+++ b/api/apps/restful_apis/dify_retrieval_api.py
@@ -276,7 +276,7 @@ async def retrieval(tenant_id):
page_size=top,
similarity_threshold=similarity_threshold,
vector_similarity_weight=0.3,
- top=top,
+ knn_top_k=top,
doc_ids=doc_ids,
rank_feature=label_question(question, [kb]),
)
diff --git a/api/apps/services/dataset_api_service.py b/api/apps/services/dataset_api_service.py
index aa6ba965e1..842e78d126 100644
--- a/api/apps/services/dataset_api_service.py
+++ b/api/apps/services/dataset_api_service.py
@@ -1061,7 +1061,8 @@ async def search(dataset_id: str, tenant_id: str, req: dict):
use_kg = req.get("use_kg", False)
similarity_threshold = float(req.get("similarity_threshold", 0.0))
vector_similarity_weight = float(req.get("vector_similarity_weight", 0.3))
- top = max(1, min(int(req.get("top_k", 1024)), 2048))
+ knn_top_k = max(1, min(int(req.get("knn_top_k", 1024)), 2048))
+ knn_num_candidates = int(req.get("knn_num_candidates", 2048))
langs = req.get("cross_languages", [])
if not KnowledgebaseService.accessible(dataset_id, tenant_id):
@@ -1090,18 +1091,18 @@ async def search(dataset_id: str, tenant_id: str, req: dict):
meta_data_filter = search_config.get("meta_data_filter", {})
similarity_threshold = float(search_config.get("similarity_threshold", similarity_threshold))
vector_similarity_weight = float(search_config.get("vector_similarity_weight", vector_similarity_weight))
- top = max(1, min(int(search_config.get("top_k", top)), 2048))
+ knn_top_k = max(1, min(int(search_config.get("top_k", knn_top_k)), 2048))
rerank_candidates_count = int(search_config.get("rerank_candidates_count", 100))
use_kg = search_config.get("use_kg", use_kg)
langs = search_config.get("cross_languages", langs)
logging.debug(
- "Dataset search loaded Search config: search_id=%s dataset_id=%s vector_similarity_weight=%s full_text_weight=%s similarity_threshold=%s top_k=%s",
+ "Dataset search loaded Search config: search_id=%s dataset_id=%s vector_similarity_weight=%s full_text_weight=%s similarity_threshold=%s knn_top_k=%s",
search_id,
dataset_id,
vector_similarity_weight,
1 - vector_similarity_weight,
similarity_threshold,
- top,
+ knn_top_k,
)
if meta_data_filter.get("method") in ["auto", "semi_auto"]:
chat_id = search_config.get("chat_id", "")
@@ -1167,7 +1168,8 @@ async def search(dataset_id: str, tenant_id: str, req: dict):
similarity_threshold,
vector_similarity_weight,
doc_ids=local_doc_ids,
- top=top,
+ knn_top_k=knn_top_k,
+ knn_num_candidates=knn_num_candidates,
rerank_mdl=rerank_mdl,
rank_feature=labels,
trace_id=search_id,
@@ -1444,7 +1446,8 @@ async def search_datasets(tenant_id: str, req: dict):
use_kg = req.get("use_kg", False)
similarity_threshold = float(req.get("similarity_threshold", 0.0))
vector_similarity_weight = float(req.get("vector_similarity_weight", 0.3))
- top = max(1, min(int(req.get("top_k", 1024)), 2048))
+ knn_top_k = max(1, min(int(req.get("knn_top_k", 1024)), 2048))
+ knn_num_candidates = int(req.get("knn_num_candidates", 2048))
langs = req.get("cross_languages", [])
logging.debug(
@@ -1485,18 +1488,18 @@ async def search_datasets(tenant_id: str, req: dict):
meta_data_filter = search_config.get("meta_data_filter", {})
similarity_threshold = float(search_config.get("similarity_threshold", similarity_threshold))
vector_similarity_weight = float(search_config.get("vector_similarity_weight", vector_similarity_weight))
- top = max(1, min(int(search_config.get("top_k", top)), 2048))
+ knn_top_k = max(1, min(int(search_config.get("top_k", knn_top_k)), 2048))
rerank_candidates_count = int(search_config.get("rerank_candidates_count", 100))
use_kg = search_config.get("use_kg", use_kg)
langs = search_config.get("cross_languages", langs)
logging.debug(
- "Dataset search loaded Search config: search_id=%s dataset_ids=%s vector_similarity_weight=%s full_text_weight=%s similarity_threshold=%s top_k=%s",
+ "Dataset search loaded Search config: search_id=%s dataset_ids=%s vector_similarity_weight=%s full_text_weight=%s similarity_threshold=%s knn_top_k=%s",
search_id,
kb_ids,
vector_similarity_weight,
1 - vector_similarity_weight,
similarity_threshold,
- top,
+ knn_top_k,
)
if meta_data_filter.get("method") in ["auto", "semi_auto"]:
chat_id = search_config.get("chat_id", "")
@@ -1566,7 +1569,8 @@ async def search_datasets(tenant_id: str, req: dict):
similarity_threshold,
vector_similarity_weight,
doc_ids=local_doc_ids,
- top=top,
+ knn_top_k=knn_top_k,
+ knn_num_candidates=knn_num_candidates,
rerank_mdl=rerank_mdl,
rank_feature=labels,
trace_id=search_id,
@@ -3875,7 +3879,7 @@ async def _search_layers_chunks(tenant_id, dataset_id, query, top_k, embd_mdl, k
kwargs = {}
if top_k is not None:
- kwargs["top"] = top_k
+ kwargs["knn_top_k"] = top_k
if doc_scope:
kwargs["doc_ids"] = [str(d) for d in doc_scope if str(d).strip()]
diff --git a/api/db/services/dialog_service.py b/api/db/services/dialog_service.py
index 1ad91c3408..53a6223986 100644
--- a/api/db/services/dialog_service.py
+++ b/api/db/services/dialog_service.py
@@ -780,7 +780,7 @@ async def async_chat(dialog, messages, stream=True, **kwargs):
dialog.similarity_threshold,
dialog.vector_similarity_weight,
doc_ids=scoped_doc_ids,
- top=dialog.top_k,
+ knn_top_k=dialog.top_k,
aggs=True,
rerank_mdl=rerank_mdl,
rank_feature=label_question(" ".join(questions), kbs),
@@ -1827,7 +1827,7 @@ async def async_ask(question, kb_ids, tenant_id, chat_llm_name=None, search_conf
page_size=12,
similarity_threshold=search_config.get("similarity_threshold", 0.1),
vector_similarity_weight=vector_similarity_weight,
- top=search_config.get("top_k", 1024),
+ knn_top_k=search_config.get("top_k", 1024),
doc_ids=doc_ids,
aggs=True,
rerank_mdl=rerank_mdl,
@@ -1930,7 +1930,7 @@ async def gen_mindmap(question, kb_ids, tenant_id, search_config={}):
page_size=12,
similarity_threshold=search_config.get("similarity_threshold", 0.2),
vector_similarity_weight=search_config.get("vector_similarity_weight", 0.3),
- top=search_config.get("top_k", 1024),
+ knn_top_k=search_config.get("top_k", 1024),
doc_ids=doc_ids,
aggs=False,
rerank_mdl=rerank_mdl,
diff --git a/api/utils/validation_utils.py b/api/utils/validation_utils.py
index 0993b8cde0..6ecdc47371 100644
--- a/api/utils/validation_utils.py
+++ b/api/utils/validation_utils.py
@@ -24,7 +24,7 @@ from typing import Annotated, Any, Literal, Union, get_args, get_origin
from uuid import UUID
from quart import Request
-from pydantic import BaseModel, ConfigDict, Field, StringConstraints, ValidationError, field_validator, model_validator, ValidationInfo
+from pydantic import AliasChoices, BaseModel, ConfigDict, Field, StringConstraints, ValidationError, field_validator, model_validator, ValidationInfo
from pydantic_core import PydanticCustomError
from werkzeug.exceptions import BadRequest, UnsupportedMediaType
@@ -982,7 +982,8 @@ class SearchDatasetReq(BaseModel):
doc_ids: Annotated[list[str], Field(default=[])]
page: Annotated[int, Field(default=1, ge=1)]
size: Annotated[int, Field(default=30, ge=1)]
- top_k: Annotated[int, Field(default=1024, ge=1)]
+ knn_top_k: Annotated[int, Field(default=1024, ge=1, le=2048, validation_alias=AliasChoices("knn_top_k", "top_k"))]
+ knn_num_candidates: Annotated[int, Field(default=2048, ge=1, le=10000)]
similarity_threshold: Annotated[float, Field(default=0.0, ge=0.0, le=1.0)]
vector_similarity_weight: Annotated[float, Field(default=0.3, ge=0.0, le=1.0)]
use_kg: Annotated[bool, Field(default=False)]
@@ -994,6 +995,12 @@ class SearchDatasetReq(BaseModel):
meta_data_filter: Annotated[dict | None, Field(default=None)]
include_knowledge_compilation: Annotated[bool, Field(default=True)]
+ @model_validator(mode="after")
+ def validate_knn_parameters(self):
+ if self.knn_num_candidates < self.knn_top_k:
+ raise ValueError("knn_num_candidates must be greater than or equal to knn_top_k")
+ return self
+
class SearchDatasetsReq(BaseModel):
"""Request model for searching multiple datasets."""
@@ -1005,7 +1012,8 @@ class SearchDatasetsReq(BaseModel):
doc_ids: Annotated[list[str], Field(default=[])]
page: Annotated[int, Field(default=1, ge=1)]
size: Annotated[int, Field(default=30, ge=1)]
- top_k: Annotated[int, Field(default=1024, ge=1)]
+ knn_top_k: Annotated[int, Field(default=1024, ge=1, le=2048, validation_alias=AliasChoices("knn_top_k", "top_k"))]
+ knn_num_candidates: Annotated[int, Field(default=2048, ge=1, le=10000)]
similarity_threshold: Annotated[float, Field(default=0.0, ge=0.0, le=1.0)]
vector_similarity_weight: Annotated[float, Field(default=0.3, ge=0.0, le=1.0)]
use_kg: Annotated[bool, Field(default=False)]
@@ -1017,6 +1025,12 @@ class SearchDatasetsReq(BaseModel):
meta_data_filter: Annotated[dict | None, Field(default=None)]
include_knowledge_compilation: Annotated[bool, Field(default=True)]
+ @model_validator(mode="after")
+ def validate_knn_parameters(self):
+ if self.knn_num_candidates < self.knn_top_k:
+ raise ValueError("knn_num_candidates must be greater than or equal to knn_top_k")
+ return self
+
class BaseListReq(BaseModel):
"""Shared pagination and sorting fields for list APIs."""
diff --git a/common/doc_store/doc_store_base.py b/common/doc_store/doc_store_base.py
index edecd0531c..6e5bdfaee8 100644
--- a/common/doc_store/doc_store_base.py
+++ b/common/doc_store/doc_store_base.py
@@ -65,7 +65,7 @@ class MatchTextExpr:
):
self.fields = fields
self.matching_text = matching_text
- self.topn = topn
+ self.topn = topn # Not used in ES, but used in Infinity
self.extra_options = extra_options
diff --git a/rag/advanced_rag/harness/tools/exploration.py b/rag/advanced_rag/harness/tools/exploration.py
index 3cc75763e8..3b724c0946 100644
--- a/rag/advanced_rag/harness/tools/exploration.py
+++ b/rag/advanced_rag/harness/tools/exploration.py
@@ -63,7 +63,7 @@ async def wiki_query(tools, query: str, keywords: str = "") -> dict:
exprs = [match_text]
if getattr(tools, "embed_mdl", None):
try:
- match_dense = await settings.retriever.get_vector(text, tools.embed_mdl, _WIKI_QUERY_TOP_N, 0.1)
+ match_dense = await settings.retriever.get_vector(text, tools.embed_mdl, top_k=_WIKI_QUERY_TOP_N, num_candidates=_WIKI_QUERY_TOP_N * 2, similarity=0.1)
exprs = [match_text, match_dense, FusionExpr("weighted_sum", _WIKI_QUERY_TOP_N, {"weights": "0.001, 1"})]
except Exception:
_LOG.exception("[Wiki lookup] dense expr build failed; BM25 only")
diff --git a/rag/advanced_rag/harness/tools/navigation.py b/rag/advanced_rag/harness/tools/navigation.py
index 7dcbec2942..674ea6d6cb 100644
--- a/rag/advanced_rag/harness/tools/navigation.py
+++ b/rag/advanced_rag/harness/tools/navigation.py
@@ -778,7 +778,7 @@ async def _kg_search(
knn_topn = pool or top_n
if getattr(tools, "embed_mdl", None):
try:
- exprs.append(await settings.retriever.get_vector(text, tools.embed_mdl, knn_topn, similarity))
+ exprs.append(await settings.retriever.get_vector(text, tools.embed_mdl, top_k=knn_topn, num_candidates=knn_topn * 2, similarity=similarity))
except Exception:
_LOG.exception("[Graph exploration] vector build failed; using keyword match")
if not exprs:
diff --git a/rag/advanced_rag/harness/tools/search.py b/rag/advanced_rag/harness/tools/search.py
index 2753a50bec..1307f5f67a 100644
--- a/rag/advanced_rag/harness/tools/search.py
+++ b/rag/advanced_rag/harness/tools/search.py
@@ -437,7 +437,7 @@ async def _search_compiled_rows(
embd_mdl = getattr(tools, "embed_mdl", None)
if embd_mdl:
try:
- exprs.append(await settings.retriever.get_vector(text, embd_mdl, top_n, 0.1))
+ exprs.append(await settings.retriever.get_vector(text, embd_mdl, top_k=top_n, num_candidates=top_n * 2, similarity=0.1))
except Exception:
_LOG.exception("[Compiled expand] vector build failed; using keyword match")
if not exprs:
@@ -681,7 +681,7 @@ async def _search_synthesis_pages(
embd_mdl = getattr(tools, "embed_mdl", None)
if embd_mdl:
try:
- exprs.append(await settings.retriever.get_vector(text, embd_mdl, top_n, 0.1))
+ exprs.append(await settings.retriever.get_vector(text, embd_mdl, top_k=top_n, num_candidates=top_n * 2, similarity=0.1))
except Exception:
_LOG.exception("[Wiki expand] vector build failed; using keyword match")
if not exprs:
diff --git a/rag/graphrag/search.py b/rag/graphrag/search.py
index d5a5efaf62..b46f295bf6 100644
--- a/rag/graphrag/search.py
+++ b/rag/graphrag/search.py
@@ -107,21 +107,21 @@ class KGSearch(Dealer):
res[(f, t)] = {"sim": get_float(ent.get("_score", 0)), "pagerank": get_float(ent.get("weight_int", 0)), "description": ent["content_with_weight"]}
return res
- def get_relevant_ents_by_keywords(self, keywords, filters, idxnms, kb_ids, emb_mdl, sim_thr=0.3, N=56):
+ async def get_relevant_ents_by_keywords(self, keywords, filters, idxnms, kb_ids, emb_mdl, sim_thr=0.3, N=56):
if not keywords:
return {}
filters = deepcopy(filters)
filters["knowledge_graph_kwd"] = "entity"
- matchDense = self.get_vector(", ".join(keywords), emb_mdl, 1024, sim_thr)
+ matchDense = await self.get_vector(", ".join(keywords), emb_mdl, top_k=1024, num_candidates=2048, similarity=sim_thr)
es_res = self.dataStore.search(["content_with_weight", "entity_kwd", "rank_flt", "n_hop_with_weight"], [], filters, [matchDense], OrderByExpr(), 0, N, idxnms, kb_ids)
return self._ent_info_from_(es_res, sim_thr)
- def get_relevant_relations_by_txt(self, txt, filters, idxnms, kb_ids, emb_mdl, sim_thr=0.3, N=56):
+ async def get_relevant_relations_by_txt(self, txt, filters, idxnms, kb_ids, emb_mdl, sim_thr=0.3, N=56):
if not txt:
return {}
filters = deepcopy(filters)
filters["knowledge_graph_kwd"] = "relation"
- matchDense = self.get_vector(txt, emb_mdl, 1024, sim_thr)
+ matchDense = await self.get_vector(txt, emb_mdl, top_k=1024, num_candidates=2048, similarity=sim_thr)
es_res = self.dataStore.search(["content_with_weight", "_score", "from_entity_kwd", "to_entity_kwd", "weight_int"], [], filters, [matchDense], OrderByExpr(), 0, N, idxnms, kb_ids)
return self._relation_info_from_(es_res, sim_thr)
@@ -165,9 +165,9 @@ class KGSearch(Dealer):
ents = [qst]
pass
- ents_from_query = self.get_relevant_ents_by_keywords(ents, filters, idxnms, kb_ids, emb_mdl, ent_sim_threshold)
+ ents_from_query = await self.get_relevant_ents_by_keywords(ents, filters, idxnms, kb_ids, emb_mdl, ent_sim_threshold)
ents_from_types = self.get_relevant_ents_by_types(ty_kwds, filters, idxnms, kb_ids, 10000)
- rels_from_txt = self.get_relevant_relations_by_txt(qst, filters, idxnms, kb_ids, emb_mdl, rel_sim_threshold)
+ rels_from_txt = await self.get_relevant_relations_by_txt(qst, filters, idxnms, kb_ids, emb_mdl, rel_sim_threshold)
nhop_pathes = defaultdict(dict)
for _, ent in ents_from_query.items():
nhops = ent.get("n_hop_ents", [])
diff --git a/rag/nlp/search.py b/rag/nlp/search.py
index dd1e6a6be2..3839a38659 100644
--- a/rag/nlp/search.py
+++ b/rag/nlp/search.py
@@ -61,14 +61,14 @@ class Dealer:
keywords: list[str] | None = None
group_docs: list[list] | None = None
- async def get_vector(self, txt, emb_mdl, topk=10, similarity=0.1):
+ async def get_vector(self, txt, emb_mdl, top_k=10, num_candidates=20, similarity=0.1):
qv, _ = await thread_pool_exec(emb_mdl.encode_queries, txt)
shape = np.array(qv).shape
if len(shape) > 1:
raise Exception(f"Dealer.get_vector returned array's shape {shape} doesn't match expectation(exact one dimension).")
embedding_data = [get_float(v) for v in qv]
vector_column_name = f"q_{len(embedding_data)}_vec"
- return MatchDenseExpr(vector_column_name, embedding_data, "float", "cosine", topk, {"similarity": similarity})
+ return MatchDenseExpr(vector_column_name, embedding_data, "float", "cosine", top_k, {"similarity": similarity, "num_candidates": num_candidates})
async def _existing_doc_ids(self, doc_ids: list[str]) -> set[str]:
if not doc_ids:
@@ -148,10 +148,13 @@ class Dealer:
orderBy = OrderByExpr()
pg = int(req.get("page", 1)) - 1
- topk = int(req.get("topk", 1024))
- ps = int(req.get("size", topk))
+ # Result pagination is independent of the KNN candidate pool size.
+ ps = int(req.get("size", 30))
offset, limit = pg * ps, ps
+ knn_top_k = int(req.get("knn_top_k", 1024))
+ knn_num_candidates = int(req.get("knn_num_candidates", 2048))
+
src = req.get(
"fields",
[
@@ -205,7 +208,7 @@ class Dealer:
total = self.dataStore.get_total(res)
logging.debug("Dealer.search TOTAL: {}".format(total))
else:
- matchDense = await self.get_vector(qst, emb_mdl, topk, req.get("similarity", 0.1))
+ matchDense = await self.get_vector(qst, emb_mdl, top_k=knn_top_k, num_candidates=knn_num_candidates, similarity=req.get("similarity", 0.1))
q_vec = matchDense.embedding_data
# ES path no longer fetches chunk vectors here. The clean
# cosine score is recovered later via a second KNN-only call
@@ -219,16 +222,16 @@ class Dealer:
if settings.DOC_ENGINE_INFINITY:
vector_similarity_weight = float(req.get("vector_similarity_weight", 0.3))
logging.debug(
- "Dealer.search fusion: topk=%s vector_similarity_weight=%s",
- topk,
+ "Dealer.search fusion: knn_top_k=%s vector_similarity_weight=%s",
+ knn_top_k,
vector_similarity_weight,
)
- fusionExpr = build_fusion_expr(topk, vector_similarity_weight)
+ fusionExpr = build_fusion_expr(knn_top_k, vector_similarity_weight)
elif settings.DOC_ENGINE_GAUSSDB:
vector_weight = req.get("vector_similarity_weight", 0.3)
- fusionExpr = FusionExpr("weighted_sum", topk, {"weights": f"{1 - float(vector_weight)},{float(vector_weight)}"})
+ fusionExpr = FusionExpr("weighted_sum", knn_top_k, {"weights": f"{1 - float(vector_weight)},{float(vector_weight)}"})
else:
- fusionExpr = FusionExpr("weighted_sum", topk, {"weights": "0.001,1"})
+ fusionExpr = FusionExpr("weighted_sum", knn_top_k, {"weights": "0.001,1"})
matchExprs = [matchText, matchDense, fusionExpr] if matchText else [matchDense]
res = await thread_pool_exec(self.dataStore.search, src, highlightFields, filters, matchExprs, orderBy, offset, limit, idx_names, kb_ids, rank_feature=rank_feature)
@@ -561,7 +564,6 @@ class Dealer:
page_size, # it is topn when rerank_mdl is specified
similarity_threshold=0.2,
vector_similarity_weight=0.3,
- top=1024, # for knn, no need to let user pass in.
doc_ids=None,
aggs=True,
rerank_mdl=None,
@@ -570,6 +572,8 @@ class Dealer:
trace_id=None,
must_not: dict | None = None,
rerank_candidates_count=64,
+ knn_top_k=1024, # Advanced knn parameter
+ knn_num_candidates=2048, # Advanced knn parameter
):
"""
Pagination is neither efficient nor reliable for this retrieval when rerank is enabled because the system must:
@@ -599,10 +603,11 @@ class Dealer:
"size": rerank_candidates_count,
"question": question,
"vector": True,
- "topk": top,
"similarity": similarity_threshold,
"available_int": 1,
"vector_similarity_weight": vector_similarity_weight,
+ "knn_top_k": knn_top_k,
+ "knn_num_candidates": knn_num_candidates,
}
if isinstance(must_not, dict) and must_not:
req["must_not"] = must_not
diff --git a/rag/utils/es_conn.py b/rag/utils/es_conn.py
index 1b15722c72..1919ee7e1b 100644
--- a/rag/utils/es_conn.py
+++ b/rag/utils/es_conn.py
@@ -256,10 +256,14 @@ class ESConnection(ESConnectionBase):
if "similarity" in m.extra_options:
similarity = m.extra_options["similarity"]
k = min(m.topn, 10000)
+ if "num_candidates" in m.extra_options:
+ num_candidates = max(k, min(m.extra_options["num_candidates"], 10000))
+ else:
+ num_candidates = min(k * 2, 10000)
s = s.knn(
m.vector_column_name,
k,
- min(k * 2, 10000),
+ num_candidates,
query_vector=list(m.embedding_data),
filter=bool_query.to_dict(), # filter=_build_knn_filter_query(bool_query, vector_similarity_weight),
similarity=similarity,
diff --git a/rag/utils/infinity_conn.py b/rag/utils/infinity_conn.py
index a869a5c20d..947220bb19 100644
--- a/rag/utils/infinity_conn.py
+++ b/rag/utils/infinity_conn.py
@@ -246,6 +246,7 @@ class InfinityConnection(InfinityConnectionBase):
matchExpr.extra_options[k] = str(v)
self.logger.debug(f"INFINITY search MatchTextExpr: {json.dumps(matchExpr.__dict__)}")
elif isinstance(matchExpr, MatchDenseExpr):
+ matchExpr.extra_options.pop("num_candidates", None)
if filter_fulltext and "filter" not in matchExpr.extra_options:
matchExpr.extra_options.update({"filter": filter_fulltext})
# dense_filter = _build_dense_filter(filter_cond, filter_fulltext, vector_similarity_weight)
diff --git a/test/testcases/test_http_api/test_session_management/test_session_sdk_routes_unit.py b/test/testcases/test_http_api/test_session_management/test_session_sdk_routes_unit.py
index 70021acca6..baa62fad59 100644
--- a/test/testcases/test_http_api/test_session_management/test_session_sdk_routes_unit.py
+++ b/test/testcases/test_http_api/test_session_management/test_session_sdk_routes_unit.py
@@ -1982,7 +1982,7 @@ def test_searchbots_retrieval_test_embedded_matrix_unit(monkeypatch):
assert retrieval_capture["question"] == "translated-q-translated"
assert retrieval_capture["similarity_threshold"] == 0.42
assert retrieval_capture["vector_similarity_weight"] == 0.8
- assert retrieval_capture["top"] == 7
+ assert retrieval_capture["knn_top_k"] == 7
assert retrieval_capture["local_doc_ids"] == ["doc-filtered"]
assert retrieval_capture["rank_feature"] == ["label-1"]
assert retrieval_capture["rerank_mdl"] is not None
diff --git a/test/unit_test/api/apps/sdk/test_dify_retrieval.py b/test/unit_test/api/apps/sdk/test_dify_retrieval.py
index 358d65fe2f..d6453407f9 100644
--- a/test/unit_test/api/apps/sdk/test_dify_retrieval.py
+++ b/test/unit_test/api/apps/sdk/test_dify_retrieval.py
@@ -427,7 +427,7 @@ class TestDifyRetrievalArgumentValidation:
kwargs = module._fake_retriever.last_kwargs
assert kwargs["page_size"] == 7
- assert kwargs["top"] == 7
+ assert kwargs["knn_top_k"] == 7
assert kwargs["similarity_threshold"] == 0.3
@pytest.mark.p1
diff --git a/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py b/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py
index 67dddce3be..8a37ec4aea 100644
--- a/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py
+++ b/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py
@@ -1692,7 +1692,7 @@ def test_tc_sql_1001_use_sql_none_falls_back_to_retrieval(
0.2,
0.3,
doc_ids=None,
- top=32,
+ knn_top_k=32,
aggs=True,
rerank_mdl=None,
rank_feature=None,
@@ -1741,7 +1741,7 @@ def test_tc_sql_1002_validator_rejection_falls_back_to_retrieval(
0.2,
0.3,
doc_ids=None,
- top=32,
+ knn_top_k=32,
aggs=True,
rerank_mdl=None,
rank_feature=None,
@@ -1792,7 +1792,7 @@ def test_tc_sql_1003_sql_timeout_falls_back_to_retrieval(
0.2,
0.3,
doc_ids=None,
- top=32,
+ knn_top_k=32,
aggs=True,
rerank_mdl=None,
rank_feature=None,
diff --git a/test/unit_test/rag/nlp/test_gaussdb_retrieval.py b/test/unit_test/rag/nlp/test_gaussdb_retrieval.py
index 4032eccc5e..3a887d5cc9 100644
--- a/test/unit_test/rag/nlp/test_gaussdb_retrieval.py
+++ b/test/unit_test/rag/nlp/test_gaussdb_retrieval.py
@@ -233,8 +233,8 @@ async def test_tc_ret_801_dealer_search_uses_requested_gaussdb_fusion_weight(
dealer = make_dealer(dealer_cls)
dealer.qryr = FakeQueryer()
- async def fake_get_vector(_text, _emb_mdl, topk=10, similarity=0.1):
- return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", topk, {"similarity": similarity})
+ async def fake_get_vector(_text, _emb_mdl, top_k=10, num_candidates=20, similarity=0.1):
+ return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", top_k, {"similarity": similarity})
dealer.get_vector = fake_get_vector
@@ -243,7 +243,7 @@ async def test_tc_ret_801_dealer_search_uses_requested_gaussdb_fusion_weight(
"question": "risk",
"page": 1,
"size": 10,
- "topk": 10,
+ "knn_top_k": 10,
"similarity": 0.2,
"vector_similarity_weight": vector_weight,
},
@@ -271,8 +271,8 @@ async def test_tc_ret_802_dealer_search_uses_default_gaussdb_fusion_weight(deale
dealer = make_dealer(dealer_cls)
dealer.qryr = FakeQueryer()
- async def fake_get_vector(_text, _emb_mdl, topk=10, similarity=0.1):
- return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", topk, {"similarity": similarity})
+ async def fake_get_vector(_text, _emb_mdl, top_k=10, num_candidates=20, similarity=0.1):
+ return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", top_k, {"similarity": similarity})
dealer.get_vector = fake_get_vector
@@ -308,12 +308,12 @@ async def test_tc_ret_306_retrieval_passes_threshold_with_original_similarity_ke
page_size=10,
similarity_threshold=0.2,
vector_similarity_weight=0.75,
- top=77,
+ knn_top_k=77,
doc_ids=["doc1"],
)
assert captured["req"]["doc_ids"] == ["doc1"]
- assert captured["req"]["topk"] == 77
+ assert captured["req"]["knn_top_k"] == 77
assert captured["req"]["similarity"] == 0.2
assert captured["req"]["available_int"] == 1
assert captured["req"]["vector_similarity_weight"] == 0.75
@@ -331,8 +331,8 @@ async def _run_tc_ret_807_808_retrieval_scenario(dealer_cls):
"mid": retrieval_chunk(0.4, "doc-mid", "Mid", "mid"),
}
- async def fake_get_vector(_text, _emb_mdl, topk=10, similarity=0.1):
- return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", topk, {"similarity": similarity})
+ async def fake_get_vector(_text, _emb_mdl, top_k=10, num_candidates=20, similarity=0.1):
+ return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", top_k, {"similarity": similarity})
async def fake_prune(sres):
return sres
@@ -379,7 +379,7 @@ async def test_tc_ret_807_retrieval_uses_gaussdb_fusion_weight_without_es_defaul
"question": "risk",
"page": 1,
"size": 10,
- "topk": 10,
+ "knn_top_k": 10,
"similarity": 0.2,
"vector_similarity_weight": 0.75,
},
@@ -420,8 +420,8 @@ async def _run_gaussdb_rank_feature_retrieval(dealer_cls, fields, rank_feature):
dealer.dataStore.ids = list(fields)
dealer.dataStore.fields = fields
- async def fake_get_vector(_text, _emb_mdl, topk=10, similarity=0.1):
- return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", topk, {"similarity": similarity})
+ async def fake_get_vector(_text, _emb_mdl, top_k=10, num_candidates=20, similarity=0.1):
+ return MatchDenseExpr("q_4_vec", [0.1, 0.2, 0.3, 0.4], "float", "cosine", top_k, {"similarity": similarity})
async def fake_prune(sres):
return sres
diff --git a/test/unit_test/rag/test_search_fusion_weight.py b/test/unit_test/rag/test_search_fusion_weight.py
index fa04b203d4..9dc3ad7c98 100644
--- a/test/unit_test/rag/test_search_fusion_weight.py
+++ b/test/unit_test/rag/test_search_fusion_weight.py
@@ -132,7 +132,7 @@ async def test_dealer_retrieval_passes_vector_similarity_weight_to_fusion_expr(s
page_size=10,
similarity_threshold=0.0,
vector_similarity_weight=0.8,
- top=10,
+ knn_top_k=10,
aggs=False,
)
@@ -163,7 +163,7 @@ async def test_dealer_retrieval_keeps_elasticsearch_fusion_weight(search_environ
page_size=10,
similarity_threshold=0.0,
vector_similarity_weight=0.8,
- top=10,
+ knn_top_k=10,
aggs=False,
)
diff --git a/web/src/components/rerank.tsx b/web/src/components/rerank.tsx
index e748524464..289c948649 100644
--- a/web/src/components/rerank.tsx
+++ b/web/src/components/rerank.tsx
@@ -27,14 +27,6 @@ import {
FormMessage,
} from './ui/form';
-export const topKSchema = {
- top_k: z.number().optional(),
-};
-
-export const initialTopKValue = {
- top_k: 1024,
-};
-
const DefaultRerankId = 'rerank_id';
interface RerankFormFieldProps {
@@ -74,7 +66,6 @@ function RerankFormField({
export const rerankFormSchema = {
[DefaultRerankId]: z.string().optional(),
- top_k: z.coerce.number().optional(),
};
interface RerankFormFieldsProps {
diff --git a/web/src/interfaces/request/knowledge.ts b/web/src/interfaces/request/knowledge.ts
index 498fc54328..d729702e4f 100644
--- a/web/src/interfaces/request/knowledge.ts
+++ b/web/src/interfaces/request/knowledge.ts
@@ -5,7 +5,6 @@ export interface ITestRetrievalRequestBody {
similarity_threshold: number;
vector_similarity_weight: number;
rerank_id?: string;
- top_k?: number;
use_kg?: boolean;
highlight?: boolean;
kb_id?: string[];
diff --git a/web/src/pages/agent/constant/index.tsx b/web/src/pages/agent/constant/index.tsx
index b0c1a863c0..ac50eaa1aa 100644
--- a/web/src/pages/agent/constant/index.tsx
+++ b/web/src/pages/agent/constant/index.tsx
@@ -88,7 +88,6 @@ export const initialRetrievalValues = {
query: AgentGlobalsSysQueryWithBrace,
top_n: 8,
rerank_candidates_count: 64,
- top_k: 1024,
kb_ids: [],
rerank_id: '',
empty_response: '',
diff --git a/web/src/pages/agent/form/retrieval-form/next.tsx b/web/src/pages/agent/form/retrieval-form/next.tsx
index 3e64f191d9..b9474189a4 100644
--- a/web/src/pages/agent/form/retrieval-form/next.tsx
+++ b/web/src/pages/agent/form/retrieval-form/next.tsx
@@ -55,7 +55,6 @@ export const RetrievalPartialSchema = {
keywords_similarity_weight: z.coerce.number().min(0).max(1),
top_n: z.coerce.number(),
...rerankCandidatesCountSchema,
- top_k: z.coerce.number(),
dataset_ids: z.array(z.string()),
rerank_id: z.string(),
empty_response: z.string(),
diff --git a/web/src/pages/agent/form/retrieval-form/use-values.ts b/web/src/pages/agent/form/retrieval-form/use-values.ts
index e6c0893fc8..5442367011 100644
--- a/web/src/pages/agent/form/retrieval-form/use-values.ts
+++ b/web/src/pages/agent/form/retrieval-form/use-values.ts
@@ -1,5 +1,5 @@
import { RAGFlowNodeType } from '@/interfaces/database/agent';
-import { isEmpty } from 'lodash';
+import { isEmpty, omit } from 'lodash';
import { useMemo } from 'react';
import { initialRetrievalValues } from '../../constant';
@@ -18,7 +18,7 @@ export function useValues(node?: RAGFlowNodeType) {
return defaultValues;
}
- return formData;
+ return omit(formData, 'top_k');
}, [defaultValues, node?.data?.form]);
return values;
diff --git a/web/src/pages/agent/form/tool-form/retrieval-form/index.tsx b/web/src/pages/agent/form/tool-form/retrieval-form/index.tsx
index bc3c6a9dc4..74115d01f1 100644
--- a/web/src/pages/agent/form/tool-form/retrieval-form/index.tsx
+++ b/web/src/pages/agent/form/tool-form/retrieval-form/index.tsx
@@ -15,6 +15,7 @@ import {
import { zodResolver } from '@hookform/resolvers/zod';
import { t } from 'i18next';
+import { omit } from 'lodash';
import { useForm } from 'react-hook-form';
import { z } from 'zod';
import { useOwnerTenantId } from '../../../context';
@@ -35,7 +36,7 @@ export const FormSchema = z.object({
});
const RetrievalForm = () => {
- const defaultValues = useValues();
+ const defaultValues = omit(useValues(), 'top_k');
const { formSchema, datasetsFetched } = useStaleDatasetFormSchema(
FormSchema,
diff --git a/web/src/pages/dataset/testing/testing-form.tsx b/web/src/pages/dataset/testing/testing-form.tsx
index f4fbf53afe..7702809699 100644
--- a/web/src/pages/dataset/testing/testing-form.tsx
+++ b/web/src/pages/dataset/testing/testing-form.tsx
@@ -14,11 +14,7 @@ import {
MetadataFilter,
MetadataFilterSchema,
} from '@/components/metadata-filter';
-import {
- RerankFormFields,
- initialTopKValue,
- topKSchema,
-} from '@/components/rerank';
+import { RerankFormFields } from '@/components/rerank';
import {
SimilaritySliderFormField,
initialSimilarityThresholdValue,
@@ -67,7 +63,6 @@ export default function TestingForm({
}),
...similarityThresholdSchema,
...vectorSimilarityWeightSchema,
- ...topKSchema,
dataset_ids: z.array(z.string()).optional(),
...MetadataFilterSchema,
size: z.number().int().min(1).max(100),
@@ -83,7 +78,6 @@ export default function TestingForm({
defaultValues: {
...initialSimilarityThresholdValue,
...initialVectorSimilarityWeightValue,
- ...initialTopKValue,
dataset_ids: [knowledgeBaseId],
size: 10,
rerank_candidates_count: 64,
diff --git a/web/src/pages/next-chats/chat/app-settings/chat-settings.tsx b/web/src/pages/next-chats/chat/app-settings/chat-settings.tsx
index 9e2315a16b..308b2b9dca 100644
--- a/web/src/pages/next-chats/chat/app-settings/chat-settings.tsx
+++ b/web/src/pages/next-chats/chat/app-settings/chat-settings.tsx
@@ -75,7 +75,6 @@ export function ChatSettings({ hasSingleChatBox }: ChatSettingsProps) {
rerank_candidates_count: 64,
similarity_threshold: 0.2,
vector_similarity_weight: 0.2,
- top_k: 1024,
meta_data_filter: {
method: DatasetMetadata.Disabled,
manual: [],
@@ -119,6 +118,7 @@ export function ChatSettings({ hasSingleChatBox }: ChatSettingsProps) {
'update_time',
'update_date',
'id',
+ 'top_k',
]),
...nextValues,
},
@@ -142,7 +142,7 @@ export function ChatSettings({ hasSingleChatBox }: ChatSettingsProps) {
: referenceMetadata;
const nextData = {
- ...data,
+ ...omit(data, 'top_k'),
prompt_config: {
...data.prompt_config,
web_search_provider: getWebSearchProvider(data.prompt_config),
diff --git a/web/src/pages/next-chats/hooks/use-create-chat.ts b/web/src/pages/next-chats/hooks/use-create-chat.ts
index 0781ab68b2..f6ae79b1f4 100644
--- a/web/src/pages/next-chats/hooks/use-create-chat.ts
+++ b/web/src/pages/next-chats/hooks/use-create-chat.ts
@@ -45,7 +45,6 @@ export const useCreateChatDialog = () => {
vector_similarity_weight: 0.3,
top_n: 8,
rerank_candidates_count: 64,
- top_k: 1024,
}),
[t, defaultModelDictionary?.llm_id],
);
diff --git a/web/src/pages/next-search/search-setting-hooks.ts b/web/src/pages/next-search/search-setting-hooks.ts
index 5435ceed0d..1fd8ef5709 100644
--- a/web/src/pages/next-search/search-setting-hooks.ts
+++ b/web/src/pages/next-search/search-setting-hooks.ts
@@ -44,7 +44,6 @@ export const SearchSettingFormSchema = z
use_kg: z.boolean(),
rerank_id: z.string(),
use_rerank: z.boolean(),
- top_k: z.number(),
...rerankCandidatesCountSchema,
summary: z.boolean(),
llm_setting: z.object({ ...LlmSettingFieldSchema, ...LLMIdFormField }),
diff --git a/web/src/pages/next-search/search-setting.tsx b/web/src/pages/next-search/search-setting.tsx
index c505c6cdf2..8ed55d3c31 100644
--- a/web/src/pages/next-search/search-setting.tsx
+++ b/web/src/pages/next-search/search-setting.tsx
@@ -43,7 +43,7 @@ import { useFetchTenantInfo } from '@/hooks/use-user-setting-request';
import { cn } from '@/lib/utils';
import { zodResolver } from '@hookform/resolvers/zod';
import { X } from 'lucide-react';
-import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { useCallback, useEffect, useMemo, useState } from 'react';
import { useForm, useWatch } from 'react-hook-form';
import { useTranslation } from 'react-i18next';
import {
@@ -110,7 +110,6 @@ function SearchSetting({
use_kg: false,
rerank_id: search_config?.rerank_id || '',
use_rerank: search_config?.rerank_id ? true : false,
- top_k: search_config?.top_k || 1024,
rerank_candidates_count: search_config?.rerank_candidates_count ?? 100,
summary: search_config?.summary || false,
chat_id: search_config?.chat_id || '',
@@ -216,15 +215,6 @@ function SearchSetting({
formMethods,
]);
- // Reset top_k to 1024 only when user actively disables rerank (from true to false)
- const prevRerankEnabled = useRef<boolean | undefined>(undefined);
- useEffect(() => {
- if (prevRerankEnabled.current === true && rerankModelEnabled === false) {
- formMethods.setValue('search_config.top_k', 1024);
- }
- prevRerankEnabled.current = rerankModelEnabled;
- }, [rerankModelEnabled, formMethods]);
-
const { updateSearch } = useUpdateSearch();
const [formSubmitLoading, setFormSubmitLoading] = useState(false);
const { data: systemSetting } = useFetchTenantInfo();