Fix: duplicated role... (#9622)

### What problem does this PR solve?

#9611
#9603 #9597

### Type of change

- [x] Bug Fix (non-breaking change which fixes an issue)

agent/canvas.py

@@ -426,7 +426,7 @@ class Canvas:
         convs = []
         if window_size <= 0:
             return convs
-        for role, obj in self.history[window_size * -1:]:
+        for role, obj in self.history[window_size * -2:]:
             if isinstance(obj, dict):
                 convs.append({"role": role, "content": obj.get("content", "")})
             else:

agent/component/base.py

@@ -36,7 +36,7 @@ _IS_RAW_CONF = "_is_raw_conf"
 
 class ComponentParamBase(ABC):
     def __init__(self):
-        self.message_history_window_size = 22
+        self.message_history_window_size = 13
         self.inputs = {}
         self.outputs = {}
         self.description = ""

agent/component/llm.py

@@ -18,11 +18,8 @@ import logging
 import os
 import re
 from typing import Any, Generator
-
 import json_repair
-from copy import deepcopy
 from functools import partial
-
 from api.db import LLMType
 from api.db.services.llm_service import LLMBundle
 from api.db.services.tenant_llm_service import TenantLLMService
@@ -130,7 +127,7 @@ class LLM(ComponentBase):
 
         args = {}
         vars = self.get_input_elements() if not self._param.debug_inputs else self._param.debug_inputs
-        prompt = self._param.sys_prompt
+        sys_prompt = self._param.sys_prompt
         for k, o in vars.items():
             args[k] = o["value"]
             if not isinstance(args[k], str):
@@ -141,14 +138,18 @@ class LLM(ComponentBase):
             self.set_input_value(k, args[k])
 
         msg = self._canvas.get_history(self._param.message_history_window_size)[:-1]
-        msg.extend(deepcopy(self._param.prompts))
-        prompt = self.string_format(prompt, args)
+        for p in self._param.prompts:
+            if msg and msg[-1]["role"] == p["role"]:
+                continue
+            msg.append(p)
+
+        sys_prompt = self.string_format(sys_prompt, args)
         for m in msg:
             m["content"] = self.string_format(m["content"], args)
         if self._param.cite and self._canvas.get_reference()["chunks"]:
-            prompt += citation_prompt()
+            sys_prompt += citation_prompt()
 
-        return prompt, msg
+        return sys_prompt, msg
 
     def _generate(self, msg:list[dict], **kwargs) -> str:
         if not self.imgs:

api/apps/sdk/dify_retrieval.py

@@ -44,9 +44,6 @@ def retrieval(tenant_id):
         if not e:
             return build_error_result(message="Knowledgebase not found!", code=settings.RetCode.NOT_FOUND)
 
-        if kb.tenant_id != tenant_id:
-            return build_error_result(message="Knowledgebase not found!", code=settings.RetCode.NOT_FOUND)
-
         embd_mdl = LLMBundle(kb.tenant_id, LLMType.EMBEDDING.value, llm_name=kb.embd_id)
 
         ranks = settings.retrievaler.retrieval(

graphrag/entity_resolution.py

@@ -169,7 +169,7 @@ class EntityResolution(Extractor):
         logging.info(f"Created resolution prompt {len(text)} bytes for {len(candidate_resolution_i[1])} entity pairs of type {candidate_resolution_i[0]}")
         async with chat_limiter:
             try:
-                with trio.move_on_after(240) as cancel_scope:
+                with trio.move_on_after(280) as cancel_scope:
                     response = await trio.to_thread.run_sync(self._chat, text, [{"role": "user", "content": "Output:"}], {})
                 if cancel_scope.cancelled_caught:
                     logging.warning("_resolve_candidate._chat timeout, skipping...")

graphrag/general/extractor.py

@@ -47,7 +47,7 @@ class Extractor:
         self._language = language
         self._entity_types = entity_types or DEFAULT_ENTITY_TYPES
 
-    @timeout(60*5)
+    @timeout(60*20)
     def _chat(self, system, history, gen_conf={}):
         hist = deepcopy(history)
         conf = deepcopy(gen_conf)

rag/raptor.py

@@ -42,7 +42,7 @@ class RecursiveAbstractiveProcessing4TreeOrganizedRetrieval:
         self._prompt = prompt
         self._max_token = max_token
 
-    @timeout(60*3)
+    @timeout(60*20)
     async def _chat(self, system, history, gen_conf):
         response = get_llm_cache(self._llm_model.llm_name, system, history, gen_conf)
         if response:
@@ -56,7 +56,7 @@ class RecursiveAbstractiveProcessing4TreeOrganizedRetrieval:
         set_llm_cache(self._llm_model.llm_name, system, response, history, gen_conf)
         return response
 
-    @timeout(2)
+    @timeout(20)
     async def _embedding_encode(self, txt):
         response = get_embed_cache(self._embd_model.llm_name, txt)
         if response is not None:
@@ -86,7 +86,7 @@ class RecursiveAbstractiveProcessing4TreeOrganizedRetrieval:
         layers = [(0, len(chunks))]
         start, end = 0, len(chunks)
 
-        @timeout(60*3)
+        @timeout(60*20)
         async def summarize(ck_idx: list[int]):
             nonlocal chunks
             texts = [chunks[i][0] for i in ck_idx]