Skip to content

armgit / dify

Public Imported from GitHub
0

Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace. Deploy on cloud, VPC, or self-hosted, so teams move from prototype to production without rebuilding the stack.

perf(api): skip saved prompts when loading chat history (#44521)

Takuyabd1edbbe55ce
diff --git a/api/core/memory/token_buffer_memory.py b/api/core/memory/token_buffer_memory.py
index bb0d50357ff..ef1a9416803 100644
--- a/api/core/memory/token_buffer_memory.py
+++ b/api/core/memory/token_buffer_memory.py
@@ -3,7 +3,7 @@ from collections.abc import Sequence
from dataclasses import dataclass
from sqlalchemy import select
-from sqlalchemy.orm import Session, sessionmaker
+from sqlalchemy.orm import Session, load_only, sessionmaker
from core.app.app_config.features.file_upload.manager import FileUploadConfigManager
from core.app.file_access import DatabaseFileAccessController
@@ -215,7 +215,24 @@ class TokenBufferMemory:
require the app. An omitted workflow repository uses the configured factory.
"""
# fetch limited messages, and return reversed
- stmt = select(Message).where(Message.conversation_id == conversation.id).order_by(Message.created_at.desc())
+ # Load only the columns used to rebuild history. Message.message stores the full prompt of each turn,
+ # earlier history included, so loading it makes memory grow quadratically with conversation length.
+ stmt = (
+ select(Message)
+ .options(
+ load_only(
+ Message.id,
+ Message.parent_message_id,
+ Message.query,
+ Message.answer,
+ Message.answer_tokens,
+ Message.workflow_run_id,
+ Message.created_at,
+ )
+ )
+ .where(Message.conversation_id == conversation.id)
+ .order_by(Message.created_at.desc())
+ )
message_limit = min(message_limit, 500) if message_limit and message_limit > 0 else 500
diff --git a/api/core/prompt/utils/get_thread_messages_length.py b/api/core/prompt/utils/get_thread_messages_length.py
index 6e4b3b1e550..6052c841342 100644
--- a/api/core/prompt/utils/get_thread_messages_length.py
+++ b/api/core/prompt/utils/get_thread_messages_length.py
@@ -1,5 +1,5 @@
from sqlalchemy import select
-from sqlalchemy.orm import Session
+from sqlalchemy.orm import Session, load_only
from core.prompt.utils.extract_thread_messages import extract_thread_messages
from models.model import Message
@@ -9,8 +9,14 @@ def get_thread_messages_length(conversation_id: str, *, session: Session) -> int
"""
Get the number of thread messages based on the parent message id.
"""
- # Fetch all messages related to the conversation
- stmt = select(Message).where(Message.conversation_id == conversation_id).order_by(Message.created_at.desc())
+ # Fetch all messages related to the conversation, loading only the columns needed to walk the thread.
+ # Message.message stores the full prompt of each turn, so loading it grows quadratically with the conversation.
+ stmt = (
+ select(Message)
+ .options(load_only(Message.id, Message.parent_message_id, Message.answer))
+ .where(Message.conversation_id == conversation_id)
+ .order_by(Message.created_at.desc())
+ )
messages = session.scalars(stmt).all()
diff --git a/api/tests/unit_tests/core/memory/test_token_buffer_memory.py b/api/tests/unit_tests/core/memory/test_token_buffer_memory.py
index 39c4ac7f14b..7a11d716b5f 100644
--- a/api/tests/unit_tests/core/memory/test_token_buffer_memory.py
+++ b/api/tests/unit_tests/core/memory/test_token_buffer_memory.py
@@ -1,6 +1,7 @@
"""Comprehensive SQLite-backed tests for token-buffer memory."""
import json
+import re
from collections.abc import Iterator, Mapping, Sequence
from dataclasses import dataclass
from datetime import UTC, datetime, timedelta
@@ -146,6 +147,7 @@ def _persist_message(
answer_tokens: int = 5,
created_at: datetime | None = None,
workflow_run_id: str | None = None,
+ parent_message_id: str | None = None,
) -> Message:
message = Message(
id=str(uuid4()),
@@ -154,6 +156,7 @@ def _persist_message(
_inputs={},
query=query,
message={},
+ parent_message_id=parent_message_id,
message_unit_price=Decimal(0),
answer=answer,
answer_tokens=answer_tokens,
@@ -375,6 +378,38 @@ class TestGetHistoryPromptMessages:
assert sum("FROM apps" in sql for sql in selects) == 1
assert len(result) == 10
+ def test_history_query_skips_saved_prompt_column(self, database: Database) -> None:
+ """Message.message stores each turn's full prompt, so history must neither select nor lazy-load it."""
+ mem = self._make_memory(database)
+ base_time = datetime.now(UTC).replace(tzinfo=None)
+ parent_message_id: str | None = None
+ for index in range(3):
+ parent_message_id = _persist_message(
+ database,
+ mem.conversation.id,
+ query=f"query-{index}",
+ answer=f"answer-{index}",
+ created_at=base_time + timedelta(seconds=index),
+ parent_message_id=parent_message_id,
+ ).id
+ # Reload rows from the database instead of reusing fully loaded instances from the identity map.
+ database.session.expunge_all()
+ before = len(database.statements)
+
+ result = mem.get_history_prompt_messages()
+
+ message_selects = [sql for sql, _ in database.statements[before:] if "FROM messages" in sql]
+ assert len(message_selects) == 1
+ assert re.search(r"\bmessages\.message\b", message_selects[0]) is None
+ assert [prompt.content for prompt in result] == [
+ "query-0",
+ "answer-0",
+ "query-1",
+ "answer-1",
+ "query-2",
+ "answer-2",
+ ]
+
@pytest.mark.parametrize(
("token_values", "max_token_limit", "expected_length"),
[
diff --git a/api/tests/unit_tests/core/prompt/test_extract_thread_messages.py b/api/tests/unit_tests/core/prompt/test_extract_thread_messages.py
index a9438e6aa18..da6c7c64981 100644
--- a/api/tests/unit_tests/core/prompt/test_extract_thread_messages.py
+++ b/api/tests/unit_tests/core/prompt/test_extract_thread_messages.py
@@ -1,8 +1,10 @@
+import re
from datetime import datetime, timedelta
from decimal import Decimal
from uuid import uuid4
import pytest
+from sqlalchemy import Engine, event
from sqlalchemy.orm import Session
from constants import UUID_NIL
@@ -198,3 +200,45 @@ def test_get_thread_messages_length_keeps_non_empty_latest_answer(sqlite_session
length = get_thread_messages_length("conversation-2", session=sqlite_session)
assert length == 2
+
+
+def test_get_thread_messages_length_skips_saved_prompt_column(sqlite_engine: Engine, sqlite_session: Session):
+ """Message.message stores each turn's full prompt, so counting the thread must neither select nor lazy-load it."""
+ id1, id2 = str(uuid4()), str(uuid4())
+ now = datetime.now()
+ sqlite_session.add_all(
+ [
+ _persisted_message(
+ message_id=id2,
+ conversation_id="conversation-3",
+ parent_message_id=id1,
+ answer="latest-answer",
+ created_at=now,
+ ),
+ _persisted_message(
+ message_id=id1,
+ conversation_id="conversation-3",
+ parent_message_id=UUID_NIL,
+ answer="older-answer",
+ created_at=now - timedelta(seconds=1),
+ ),
+ ]
+ )
+ sqlite_session.commit()
+ # Reload rows from the database instead of reusing fully loaded instances from the identity map.
+ sqlite_session.expunge_all()
+ statements: list[str] = []
+
+ def record_statement(_connection, _cursor, statement, _parameters, _context, _executemany) -> None:
+ statements.append(statement)
+
+ event.listen(sqlite_engine, "before_cursor_execute", record_statement)
+ try:
+ length = get_thread_messages_length("conversation-3", session=sqlite_session)
+ finally:
+ event.remove(sqlite_engine, "before_cursor_execute", record_statement)
+
+ assert length == 2
+ message_selects = [sql for sql in statements if "FROM messages" in sql]
+ assert len(message_selects) == 1
+ assert re.search(r"\bmessages\.message\b", message_selects[0]) is None