Fix raptor resuable issue. (#4063)

### What problem does this PR solve?

#4045

### Type of change

- [x] New Feature (non-breaking change which adds functionality)

api/db/services/document_service.py

@@ -344,6 +344,8 @@ class DocumentService(CommonService):
                     old[k] = v
 
         dfs_update(d.parser_config, config)
+        if not config.get("raptor") and d.parser_config.get("raptor"):
+            del d.parser_config["raptor"]
         cls.update_by_id(id, {"parser_config": d.parser_config})
 
     @classmethod
@@ -432,6 +434,11 @@ class DocumentService(CommonService):
 
 
 def queue_raptor_tasks(doc):
+    chunking_config = DocumentService.get_chunking_config(doc["id"])
+    hasher = xxhash.xxh64()
+    for field in sorted(chunking_config.keys()):
+        hasher.update(str(chunking_config[field]).encode("utf-8"))
+
     def new_task():
         nonlocal doc
         return {
@@ -443,6 +450,9 @@ def queue_raptor_tasks(doc):
         }
 
     task = new_task()
+    for field in ["doc_id", "from_page", "to_page"]:
+        hasher.update(str(task.get(field, "")).encode("utf-8"))
+    task["digest"] = hasher.hexdigest()
     bulk_insert_into_db(Task, [task], True)
     task["type"] = "raptor"
     assert REDIS_CONN.queue_product(SVR_QUEUE_NAME, message=task), "Can't access Redis. Please check the Redis' status."

api/db/services/task_service.py

@@ -34,15 +34,17 @@ from rag.utils.redis_conn import REDIS_CONN
 from api import settings
 from rag.nlp import search
 
+
 def trim_header_by_lines(text: str, max_length) -> str:
     len_text = len(text)
     if len_text <= max_length:
         return text
     for i in range(len_text):
         if text[i] == '\n' and len_text - i <= max_length:
-            return text[i+1:]
+            return text[i + 1:]
     return text
 
+
 class TaskService(CommonService):
     model = Task
 
@@ -212,8 +214,8 @@ def queue_tasks(doc: dict, bucket: str, name: str):
         if doc["parser_id"] == "paper":
             page_size = doc["parser_config"].get("task_page_size", 22)
         if doc["parser_id"] in ["one", "knowledge_graph"] or not do_layout:
-            page_size = 10**9
-        page_ranges = doc["parser_config"].get("pages") or [(1, 10**5)]
+            page_size = 10 ** 9
+        page_ranges = doc["parser_config"].get("pages") or [(1, 10 ** 5)]
         for s, e in page_ranges:
             s -= 1
             s = max(0, s)
@@ -257,7 +259,8 @@ def queue_tasks(doc: dict, bucket: str, name: str):
             if task["chunk_ids"]:
                 chunk_ids.extend(task["chunk_ids"].split())
         if chunk_ids:
-            settings.docStoreConn.delete({"id": chunk_ids}, search.index_name(chunking_config["tenant_id"]), chunking_config["kb_id"])
+            settings.docStoreConn.delete({"id": chunk_ids}, search.index_name(chunking_config["tenant_id"]),
+                                         chunking_config["kb_id"])
     DocumentService.update_by_id(doc["id"], {"chunk_num": ck_num})
 
     bulk_insert_into_db(Task, tsks, True)
@@ -271,7 +274,8 @@ def queue_tasks(doc: dict, bucket: str, name: str):
 
 
 def reuse_prev_task_chunks(task: dict, prev_tasks: list[dict], chunking_config: dict):
-    idx = bisect.bisect_left(prev_tasks, task.get("from_page", 0), key=lambda x: x.get("from_page",0))
+    idx = bisect.bisect_left(prev_tasks, (task.get("from_page", 0), task.get("digest", "")),
+                             key=lambda x: (x.get("from_page", 0), x.get("digest", "")))
     if idx >= len(prev_tasks):
         return 0
     prev_task = prev_tasks[idx]

graphrag/utils.py

@@ -78,7 +78,7 @@ def get_llm_cache(llmnm, txt, history, genconf):
     bin = REDIS_CONN.get(k)
     if not bin:
         return
-    return bin.decode("utf-8")
+    return bin
 
 
 def set_llm_cache(llmnm, txt, v: str, history, genconf):