pengjunjie/ragflow
1
0
Fork 1
mirror of https://github.com/infiniflow/ragflow.git synced 2025-12-08 20:42:30 +08:00
Code Issues Packages Projects Releases Wiki Activity

perf: optimze figure parser (#7392)

Browse Source 
### What problem does this PR solve?

When parsing documents containing images, the current code uses a
single-threaded approach to call the VL model, resulting in extremely
slow parsing speed (e.g., parsing a Word document with dozens of images
takes over 20 minutes).

By switching to a multithreaded approach to call the VL model, the
parsing speed can be improved to an acceptable level.

### Type of change

- [x] Performance Improvement

---------

Co-authored-by: liuzhenghua-jk <liuzhenghua-jk@360shuke.com>
This commit is contained in:
liuzhenghua
2025-05-06 14:39:45 +08:00
committed by GitHub
parent d6cc6453d1
commit 2f768b96e8
2 changed files with 43 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
rag/svr/task_executor.py
View File

@ -99,8 +99,10 @@ CURRENT_TASKS = {}
MAX_CONCURRENT_TASKS = int(os.environ.get('MAX_CONCURRENT_TASKS', "5"))
MAX_CONCURRENT_CHUNK_BUILDERS = int(os.environ.get('MAX_CONCURRENT_CHUNK_BUILDERS', "1"))
MAX_CONCURRENT_MINIO = int(os.environ.get('MAX_CONCURRENT_MINIO', '10'))
task_limiter = trio.CapacityLimiter(MAX_CONCURRENT_TASKS)
chunk_limiter = trio.CapacityLimiter(MAX_CONCURRENT_CHUNK_BUILDERS)
minio_limiter = trio.CapacityLimiter(MAX_CONCURRENT_MINIO)
WORKER_HEARTBEAT_TIMEOUT = int(os.environ.get('WORKER_HEARTBEAT_TIMEOUT', '120'))
stop_event = threading.Event()
@ -270,38 +272,43 @@ async def build_chunks(task, progress_callback):
}
if task["pagerank"]:
doc[PAGERANK_FLD] = int(task["pagerank"])
el = 0
for ck in cks:
d = copy.deepcopy(doc)
d.update(ck)
d["id"] = xxhash.xxh64((ck["content_with_weight"] + str(d["doc_id"])).encode("utf-8")).hexdigest()
d["create_time"] = str(datetime.now()).replace("T", " ")[:19]
d["create_timestamp_flt"] = datetime.now().timestamp()
if not d.get("image"):
_ = d.pop("image", None)
d["img_id"] = ""
docs.append(d)
continue
st = timer()
async def upload_to_minio(document, chunk):
try:
output_buffer = BytesIO()
if isinstance(d["image"], bytes):
output_buffer = BytesIO(d["image"])
else:
d["image"].save(output_buffer, format='JPEG')
async with minio_limiter:
d = copy.deepcopy(document)
d.update(chunk)
d["id"] = xxhash.xxh64((chunk["content_with_weight"] + str(d["doc_id"])).encode("utf-8")).hexdigest()
d["create_time"] = str(datetime.now()).replace("T", " ")[:19]
d["create_timestamp_flt"] = datetime.now().timestamp()
if not d.get("image"):
_ = d.pop("image", None)
d["img_id"] = ""
docs.append(d)
return
st = timer()
await trio.to_thread.run_sync(lambda: STORAGE_IMPL.put(task["kb_id"], d["id"], output_buffer.getvalue()))
el += timer() - st
output_buffer = BytesIO()
if isinstance(d["image"], bytes):
output_buffer = BytesIO(d["image"])
else:
d["image"].save(output_buffer, format='JPEG')
await trio.to_thread.run_sync(lambda: STORAGE_IMPL.put(task["kb_id"], d["id"], output_buffer.getvalue()))
d["img_id"] = "{}-{}".format(task["kb_id"], d["id"])
del d["image"]
docs.append(d)
except Exception:
logging.exception(
"Saving image of chunk {}/{}/{} got exception".format(task["location"], task["name"], d["id"]))
raise
d["img_id"] = "{}-{}".format(task["kb_id"], d["id"])
del d["image"]
docs.append(d)
logging.info("MINIO PUT({}):{}".format(task["name"], el))
async with trio.open_nursery() as nursery:
for ck in cks:
nursery.start_soon(upload_to_minio, doc, ck)
el = timer() - st
logging.info("MINIO PUT({}) cost {:.3f} s".format(task["name"], el))
if task["parser_config"].get("auto_keywords", 0):
st = timer()