apunkt/SurfSense
1
0
Fork 0
mirror of https://github.com/MODSetter/SurfSense.git synced 2026-06-06 20:15:17 +02:00
Code Issues Projects Releases Packages Wiki Activity

feat(backend): Remove LLM summaries from document indexing

Browse source
This commit is contained in:
Anish Sarkar 2026-06-04 00:50:19 +05:30
parent 290a9539ef
commit 81fa219b30
17 changed files with 40 additions and 518 deletions
Download patch file Download diff file Expand all files Collapse all files

82
surfsense_backend/app/tasks/document_processors/_save.py
View file

@ -1,20 +1,15 @@
"""
Unified document save/update logic for file processors.
"""
"""Unified document save/update logic for file processors."""
import asyncio
import logging
from sqlalchemy.exc import SQLAlchemyError
from sqlalchemy.ext.asyncio import AsyncSession
from app.db import Document, DocumentStatus, DocumentType
from app.services.llm_service import get_user_long_context_llm
from app.utils.document_converters import (
create_document_chunks,
embed_text,
generate_content_hash,
generate_document_summary,
)
from ._helpers import (
@ -24,59 +19,6 @@ from ._helpers import (
)
from .base import get_current_timestamp, safe_set_chunks
# ---------------------------------------------------------------------------
# Summary generation
# ---------------------------------------------------------------------------
async def _generate_summary(
markdown_content: str,
file_name: str,
etl_service: str,
user_llm,
enable_summary: bool,
) -> tuple[str, list[float]]:
"""
Generate a document summary and embedding.
Docling uses its own large-document summary strategy; other ETL services
use the standard ``generate_document_summary`` helper.
"""
if not enable_summary:
summary = f"File: {file_name}\n\n{markdown_content[:4000]}"
return summary, await asyncio.to_thread(embed_text, summary)
if etl_service == "DOCLING":
from app.services.docling_service import create_docling_service
docling_service = create_docling_service()
summary_text = await docling_service.process_large_document_summary(
content=markdown_content, llm=user_llm, document_title=file_name
)
meta = {
"file_name": file_name,
"etl_service": etl_service,
"document_type": "File Document",
}
parts = ["# DOCUMENT METADATA"]
for key, value in meta.items():
if value:
formatted_key = key.replace("_", " ").title()
parts.append(f"**{formatted_key}:** {value}")
enhanced = "\n".join(parts) + "\n\n# DOCUMENT SUMMARY\n\n" + summary_text
return enhanced, await asyncio.to_thread(embed_text, enhanced)
# Standard summary (Unstructured / LlamaCloud / others)
meta = {
"file_name": file_name,
"etl_service": etl_service,
"document_type": "File Document",
}
return await generate_document_summary(markdown_content, user_llm, meta)
# ---------------------------------------------------------------------------
# Unified save function
# ---------------------------------------------------------------------------
@ -90,7 +32,6 @@ async def save_file_document(
user_id: str,
etl_service: str,
connector: dict | None = None,
enable_summary: bool = True,
) -> Document | None:
"""
Process and store a file document with deduplication and migration support.
@ -106,7 +47,6 @@ async def save_file_document(
user_id: ID of the user
etl_service: Name of the ETL service (UNSTRUCTURED, LLAMACLOUD, DOCLING)
connector: Optional connector info for Google Drive files
enable_summary: Whether to generate an AI summary
Returns:
Document object if successful, None if duplicate detected
@ -133,24 +73,16 @@ async def save_file_document(
if should_skip:
return doc
user_llm = await get_user_long_context_llm(session, user_id, search_space_id)
if not user_llm:
raise RuntimeError(
f"No long context LLM configured for user {user_id} "
f"in search space {search_space_id}"
)
summary_content, summary_embedding = await _generate_summary(
markdown_content, file_name, etl_service, user_llm, enable_summary
)
document_content = f"File: {file_name}\n\n{markdown_content[:4000]}"
document_embedding = embed_text(document_content)
chunks = await create_document_chunks(markdown_content)
doc_metadata = {"FILE_NAME": file_name, "ETL_SERVICE": etl_service}
if existing_document:
existing_document.title = file_name
existing_document.content = summary_content
existing_document.content = document_content
existing_document.content_hash = content_hash
existing_document.embedding = summary_embedding
existing_document.embedding = document_embedding
existing_document.document_metadata = doc_metadata
await safe_set_chunks(session, existing_document, chunks)
existing_document.source_markdown = markdown_content
@ -171,8 +103,8 @@ async def save_file_document(
title=file_name,
document_type=doc_type,
document_metadata=doc_metadata,
content=summary_content,
embedding=summary_embedding,
content=document_content,
embedding=document_embedding,
chunks=chunks,
content_hash=content_hash,
unique_identifier_hash=primary_hash,

36
surfsense_backend/app/tasks/document_processors/circleback_processor.py
View file

@ -25,11 +25,10 @@ from app.db import (
SearchSourceConnectorType,
SearchSpace,
)
from app.services.llm_service import get_document_summary_llm
from app.utils.document_converters import (
create_document_chunks,
embed_text,
generate_content_hash,
generate_document_summary,
generate_unique_identifier_hash,
)
@ -176,34 +175,8 @@ async def add_circleback_meeting_document(
# PHASE 3: Process the document content
# =======================================================================
# Get LLM for generating summary
llm = await get_document_summary_llm(session, search_space_id)
if not llm:
logger.warning(
f"No LLM configured for search space {search_space_id}. Using content as summary."
)
# Use first 1000 chars as summary if no LLM available
summary_content = (
markdown_content[:1000] + "..."
if len(markdown_content) > 1000
else markdown_content
)
summary_embedding = None
else:
# Generate summary with metadata
summary_metadata = {
"meeting_name": meeting_name,
"meeting_id": meeting_id,
"document_type": "Circleback Meeting",
**{
k: v
for k, v in metadata.items()
if isinstance(v, str | int | float | bool)
},
}
summary_content, summary_embedding = await generate_document_summary(
markdown_content, llm, summary_metadata
)
summary_content = markdown_content
summary_embedding = embed_text(summary_content)
# Process chunks
chunks = await create_document_chunks(markdown_content)
@ -224,8 +197,7 @@ async def add_circleback_meeting_document(
document.title = meeting_name
document.content = summary_content
document.content_hash = content_hash
if summary_embedding is not None:
document.embedding = summary_embedding
document.embedding = summary_embedding
document.document_metadata = document_metadata
await safe_set_chunks(session, document, chunks)
document.source_markdown = markdown_content

25
surfsense_backend/app/tasks/document_processors/extension_processor.py
View file

@ -9,12 +9,11 @@ from sqlalchemy.ext.asyncio import AsyncSession
from app.db import Document, DocumentType
from app.schemas import ExtensionDocumentContent
from app.services.llm_service import get_user_long_context_llm
from app.services.task_logging_service import TaskLoggingService
from app.utils.document_converters import (
create_document_chunks,
embed_text,
generate_content_hash,
generate_document_summary,
generate_unique_identifier_hash,
)
@ -123,26 +122,8 @@ async def add_extension_received_document(
f"Content changed for URL {content.metadata.VisitedWebPageURL}. Updating document."
)
# Get user's long context LLM (needed for both create and update)
user_llm = await get_user_long_context_llm(session, user_id, search_space_id)
if not user_llm:
raise RuntimeError(
f"No long context LLM configured for user {user_id} in search space {search_space_id}"
)
# Generate summary with metadata
document_metadata = {
"session_id": content.metadata.BrowsingSessionId,
"url": content.metadata.VisitedWebPageURL,
"title": content.metadata.VisitedWebPageTitle,
"referrer": content.metadata.VisitedWebPageReffererURL,
"timestamp": content.metadata.VisitedWebPageDateWithTimeInISOString,
"duration_ms": content.metadata.VisitedWebPageVisitDurationInMilliseconds,
"document_type": "Browser Extension Capture",
}
summary_content, summary_embedding = await generate_document_summary(
combined_document_string, user_llm, document_metadata
)
summary_content = combined_document_string
summary_embedding = embed_text(summary_content)
# Process chunks
chunks = await create_document_chunks(content.pageContent)

15
surfsense_backend/app/tasks/document_processors/file_processors.py
View file

@ -10,7 +10,7 @@ from __future__ import annotations
import contextlib
import logging
import os
from dataclasses import dataclass, field
from dataclasses import dataclass
from fastapi import HTTPException
from sqlalchemy.ext.asyncio import AsyncSession
@ -48,12 +48,6 @@ class _ProcessingContext:
notification: Notification | None = None
use_vision_llm: bool = False
processing_mode: str = "basic"
enable_summary: bool = field(init=False)
def __post_init__(self) -> None:
self.enable_summary = (
self.connector.get("enable_summary", True) if self.connector else True
)
# ---------------------------------------------------------------------------
@ -261,7 +255,6 @@ async def _process_document_upload(ctx: _ProcessingContext) -> Document | None:
ctx.user_id,
etl_result.etl_service,
ctx.connector,
enable_summary=ctx.enable_summary,
)
if result:
@ -466,7 +459,6 @@ async def process_file_in_background_with_document(
log_entry: Log,
connector: dict | None = None,
notification: Notification | None = None,
should_summarize: bool = False,
use_vision_llm: bool = False,
processing_mode: str = "basic",
) -> Document | None:
@ -482,7 +474,6 @@ async def process_file_in_background_with_document(
from app.indexing_pipeline.adapters.file_upload_adapter import (
UploadDocumentAdapter,
)
from app.services.llm_service import get_user_long_context_llm
from app.utils.document_converters import generate_content_hash
from .base import check_duplicate_document
@ -522,8 +513,6 @@ async def process_file_in_background_with_document(
stage="chunking",
)
user_llm = await get_user_long_context_llm(session, user_id, search_space_id)
adapter = UploadDocumentAdapter(session)
await adapter.index(
markdown_content=markdown_content,
@ -531,8 +520,6 @@ async def process_file_in_background_with_document(
etl_service=etl_service,
search_space_id=search_space_id,
user_id=user_id,
llm=user_llm,
should_summarize=should_summarize,
)
if billable_pages > 0:

20
surfsense_backend/app/tasks/document_processors/markdown_processor.py
View file

@ -8,12 +8,11 @@ from sqlalchemy.exc import SQLAlchemyError
from sqlalchemy.ext.asyncio import AsyncSession
from app.db import Document, DocumentStatus, DocumentType
from app.services.llm_service import get_user_long_context_llm
from app.services.task_logging_service import TaskLoggingService
from app.utils.document_converters import (
create_document_chunks,
embed_text,
generate_content_hash,
generate_document_summary,
)
from ._helpers import (
@ -183,21 +182,8 @@ async def add_received_markdown_file_document(
return doc
# Content changed - continue to update
# Get user's long context LLM (needed for both create and update)
user_llm = await get_user_long_context_llm(session, user_id, search_space_id)
if not user_llm:
raise RuntimeError(
f"No long context LLM configured for user {user_id} in search space {search_space_id}"
)
# Generate summary with metadata
document_metadata = {
"file_name": file_name,
"document_type": "Markdown File Document",
}
summary_content, summary_embedding = await generate_document_summary(
file_in_markdown, user_llm, document_metadata
)
summary_content = f"File: {file_name}\n\n{file_in_markdown[:4000]}"
summary_embedding = embed_text(summary_content)
# Process chunks
chunks = await create_document_chunks(file_in_markdown)

39
surfsense_backend/app/tasks/document_processors/youtube_processor.py
View file

@ -17,12 +17,11 @@ from sqlalchemy.ext.asyncio import AsyncSession
from youtube_transcript_api import YouTubeTranscriptApi
from app.db import Document, DocumentStatus, DocumentType
from app.services.llm_service import get_user_long_context_llm
from app.services.task_logging_service import TaskLoggingService
from app.utils.document_converters import (
create_document_chunks,
embed_text,
generate_content_hash,
generate_document_summary,
generate_unique_identifier_hash,
)
from app.utils.proxy_config import get_requests_proxies
@ -355,40 +354,8 @@ async def add_youtube_video_document(
await session.commit()
return document
# Get LLM for summary generation
await task_logger.log_task_progress(
log_entry,
f"Preparing for summary generation: {video_data.get('title', 'YouTube Video')}",
{"stage": "llm_setup"},
)
# Get user's long context LLM
user_llm = await get_user_long_context_llm(session, user_id, search_space_id)
if not user_llm:
raise RuntimeError(
f"No long context LLM configured for user {user_id} in search space {search_space_id}"
)
# Generate summary
await task_logger.log_task_progress(
log_entry,
f"Generating summary for video: {video_data.get('title', 'YouTube Video')}",
{"stage": "summary_generation"},
)
# Generate summary with metadata
document_metadata_for_summary = {
"url": url,
"video_id": video_id,
"title": video_data.get("title", "YouTube Video"),
"author": video_data.get("author_name", "Unknown"),
"thumbnail": video_data.get("thumbnail_url", ""),
"document_type": "YouTube Video Document",
"has_transcript": "No captions available" not in transcript_text,
}
summary_content, summary_embedding = await generate_document_summary(
combined_document_string, user_llm, document_metadata_for_summary
)
summary_content = combined_document_string
summary_embedding = embed_text(summary_content)
# Process chunks
await task_logger.log_task_progress(