revert batch query (#17707)

This commit is contained in:
Jyong 2025-04-09 20:25:36 +08:00 committed by GitHub
parent 1d5c07dedb
commit 8b3be4224d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -1,13 +1,9 @@
import concurrent.futures import concurrent.futures
import logging
import time
from concurrent.futures import ThreadPoolExecutor from concurrent.futures import ThreadPoolExecutor
from typing import Optional from typing import Optional
from flask import Flask, current_app from flask import Flask, current_app
from sqlalchemy import and_, or_
from sqlalchemy.orm import load_only from sqlalchemy.orm import load_only
from sqlalchemy.sql.expression import false
from configs import dify_config from configs import dify_config
from core.rag.data_post_processor.data_post_processor import DataPostProcessor from core.rag.data_post_processor.data_post_processor import DataPostProcessor
@ -182,7 +178,6 @@ class RetrievalService:
if not dataset: if not dataset:
raise ValueError("dataset not found") raise ValueError("dataset not found")
start = time.time()
vector = Vector(dataset=dataset) vector = Vector(dataset=dataset)
documents = vector.search_by_vector( documents = vector.search_by_vector(
query, query,
@ -192,7 +187,6 @@ class RetrievalService:
filter={"group_id": [dataset.id]}, filter={"group_id": [dataset.id]},
document_ids_filter=document_ids_filter, document_ids_filter=document_ids_filter,
) )
logging.debug(f"embedding_search ends at {time.time() - start:.2f} seconds")
if documents: if documents:
if ( if (
@ -276,8 +270,7 @@ class RetrievalService:
return [] return []
try: try:
start_time = time.time() # Collect document IDs
# Collect document IDs with existence check
document_ids = {doc.metadata.get("document_id") for doc in documents if "document_id" in doc.metadata} document_ids = {doc.metadata.get("document_id") for doc in documents if "document_id" in doc.metadata}
if not document_ids: if not document_ids:
return [] return []
@ -295,138 +288,110 @@ class RetrievalService:
include_segment_ids = set() include_segment_ids = set()
segment_child_map = {} segment_child_map = {}
# Precompute doc_forms to avoid redundant checks # Process documents
doc_forms = {}
for doc in documents:
document_id = doc.metadata.get("document_id")
dataset_doc = dataset_documents.get(document_id)
if dataset_doc:
doc_forms[document_id] = dataset_doc.doc_form
# Batch collect index node IDs with type safety
child_index_node_ids = []
index_node_ids = []
for doc in documents:
document_id = doc.metadata.get("document_id")
if doc_forms.get(document_id) == IndexType.PARENT_CHILD_INDEX:
child_index_node_ids.append(doc.metadata.get("doc_id"))
else:
index_node_ids.append(doc.metadata.get("doc_id"))
# Batch query ChildChunk
child_chunks = db.session.query(ChildChunk).filter(ChildChunk.index_node_id.in_(child_index_node_ids)).all()
child_chunk_map = {chunk.index_node_id: chunk for chunk in child_chunks}
segment_ids_from_child = [chunk.segment_id for chunk in child_chunks]
segment_conditions = []
if index_node_ids:
segment_conditions.append(DocumentSegment.index_node_id.in_(index_node_ids))
if segment_ids_from_child:
segment_conditions.append(DocumentSegment.id.in_(segment_ids_from_child))
if segment_conditions:
filter_expr = or_(*segment_conditions)
else:
filter_expr = false()
segment_map = {
segment.id: segment
for segment in db.session.query(DocumentSegment)
.filter(
and_(
filter_expr,
DocumentSegment.enabled == True,
DocumentSegment.status == "completed",
)
)
.options(
load_only(
DocumentSegment.id,
DocumentSegment.content,
DocumentSegment.answer,
)
)
.all()
}
for document in documents: for document in documents:
document_id = document.metadata.get("document_id") document_id = document.metadata.get("document_id")
dataset_document = dataset_documents.get(document_id) if document_id not in dataset_documents:
continue
dataset_document = dataset_documents[document_id]
if not dataset_document: if not dataset_document:
continue continue
doc_form = doc_forms.get(document_id) if dataset_document.doc_form == IndexType.PARENT_CHILD_INDEX:
if doc_form == IndexType.PARENT_CHILD_INDEX: # Handle parent-child documents
# Handle parent-child documents using preloaded data
child_index_node_id = document.metadata.get("doc_id") child_index_node_id = document.metadata.get("doc_id")
if not child_index_node_id:
continue
child_chunk = child_chunk_map.get(child_index_node_id) child_chunk = (
db.session.query(ChildChunk).filter(ChildChunk.index_node_id == child_index_node_id).first()
)
if not child_chunk: if not child_chunk:
continue continue
segment = segment_map.get(child_chunk.segment_id) segment = (
db.session.query(DocumentSegment)
.filter(
DocumentSegment.dataset_id == dataset_document.dataset_id,
DocumentSegment.enabled == True,
DocumentSegment.status == "completed",
DocumentSegment.id == child_chunk.segment_id,
)
.options(
load_only(
DocumentSegment.id,
DocumentSegment.content,
DocumentSegment.answer,
)
)
.first()
)
if not segment: if not segment:
continue continue
if segment.id not in include_segment_ids: if segment.id not in include_segment_ids:
include_segment_ids.add(segment.id) include_segment_ids.add(segment.id)
map_detail = {"max_score": document.metadata.get("score", 0.0), "child_chunks": []} child_chunk_detail = {
"id": child_chunk.id,
"content": child_chunk.content,
"position": child_chunk.position,
"score": document.metadata.get("score", 0.0),
}
map_detail = {
"max_score": document.metadata.get("score", 0.0),
"child_chunks": [child_chunk_detail],
}
segment_child_map[segment.id] = map_detail segment_child_map[segment.id] = map_detail
records.append({"segment": segment}) record = {
"segment": segment,
# Append child chunk details }
child_chunk_detail = { records.append(record)
"id": child_chunk.id, else:
"content": child_chunk.content, child_chunk_detail = {
"position": child_chunk.position, "id": child_chunk.id,
"score": document.metadata.get("score", 0.0), "content": child_chunk.content,
} "position": child_chunk.position,
segment_child_map[segment.id]["child_chunks"].append(child_chunk_detail) "score": document.metadata.get("score", 0.0),
segment_child_map[segment.id]["max_score"] = max( }
segment_child_map[segment.id]["max_score"], document.metadata.get("score", 0.0) segment_child_map[segment.id]["child_chunks"].append(child_chunk_detail)
) segment_child_map[segment.id]["max_score"] = max(
segment_child_map[segment.id]["max_score"], document.metadata.get("score", 0.0)
)
else: else:
# Handle normal documents # Handle normal documents
index_node_id = document.metadata.get("doc_id") index_node_id = document.metadata.get("doc_id")
if not index_node_id: if not index_node_id:
continue continue
segment = next( segment = (
( db.session.query(DocumentSegment)
s .filter(
for s in segment_map.values() DocumentSegment.dataset_id == dataset_document.dataset_id,
if s.index_node_id == index_node_id and s.dataset_id == dataset_document.dataset_id DocumentSegment.enabled == True,
), DocumentSegment.status == "completed",
None, DocumentSegment.index_node_id == index_node_id,
)
.first()
) )
if not segment: if not segment:
continue continue
if segment.id not in include_segment_ids: include_segment_ids.add(segment.id)
include_segment_ids.add(segment.id) record = {
records.append( "segment": segment,
{ "score": document.metadata.get("score"), # type: ignore
"segment": segment, }
"score": document.metadata.get("score", 0.0), records.append(record)
}
)
# Merge child chunks information # Add child chunks information to records
for record in records: for record in records:
segment_id = record["segment"].id if record["segment"].id in segment_child_map:
if segment_id in segment_child_map: record["child_chunks"] = segment_child_map[record["segment"].id].get("child_chunks") # type: ignore
record["child_chunks"] = segment_child_map[segment_id]["child_chunks"] record["score"] = segment_child_map[record["segment"].id]["max_score"]
record["score"] = segment_child_map[segment_id]["max_score"]
logging.debug(f"Formatting retrieval documents took {time.time() - start_time:.2f} seconds")
return [RetrievalSegments(**record) for record in records] return [RetrievalSegments(**record) for record in records]
except Exception as e: except Exception as e:
# Only rollback if there were write operations
db.session.rollback() db.session.rollback()
raise e raise e