feat(zendesk): add zendesk workflow (#3586)

# Description

Fix: CLI-24

core/example_workflows/talk_to_file_rag_config_workflow.yaml

@@ -0,0 +1,39 @@
+{
+  "max_files": 20,
+  "llm_config": { "temperature": 0.3, "max_context_tokens": 20000 },
+  "max_history": 10,
+  "reranker_config":
+    { "model": "rerank-v3.5", "top_n": 10, "supplier": "cohere" },
+  "workflow_config":
+    {
+      "name": "Standard RAG",
+      "nodes":
+        [
+          {
+            "name": "START",
+            "edges": ["filter_history"],
+            "description": "Starting workflow",
+          },
+          {
+            "name": "filter_history",
+            "edges": ["retrieve"],
+            "description": "Filtering history",
+          },
+          {
+            "name": "retrieve",
+            "edges": ["retrieve_full_documents_context"],
+            "description": "Retrieving relevant information",
+          },
+          {
+            "name": "retrieve_full_documents_context",
+            "edges": ["generate_zendesk_rag"],
+            "description": "Retrieving full tickets context",
+          },
+          {
+            "name": "generate_zendesk_rag",
+            "edges": ["END"],
+            "description": "Generating answer",
+          },
+        ],
+    },
+}

core/quivr_core/rag/prompts.py

@@ -1,14 +1,14 @@
 import datetime
-from pydantic import ConfigDict, create_model
 
-from langchain_core.prompts.base import BasePromptTemplate
 from langchain_core.prompts import (
     ChatPromptTemplate,
     HumanMessagePromptTemplate,
+    MessagesPlaceholder,
     PromptTemplate,
     SystemMessagePromptTemplate,
-    MessagesPlaceholder,
 )
+from langchain_core.prompts.base import BasePromptTemplate
+from pydantic import ConfigDict, create_model
 
 
 class CustomPromptsDict(dict):
@@ -258,6 +258,32 @@ def _define_custom_prompts() -> CustomPromptsDict:
 
     custom_prompts["TOOL_ROUTING_PROMPT"] = TOOL_ROUTING_PROMPT
 
+    system_message_zendesk_template = """
+
+    - You are a Zendesk Agent.
+    - You are answering a client query.
+    - You must provide a response with all the information you have. Do not write areas to be filled like [your name], [your email], etc.
+    - Give a the most complete answer to the client query and give relevant links if needed.
+    - Based on the following similar client tickets, provide a response to the client query in the same format.
+
+    ------ Zendesk Similar Tickets ------
+    {similar_tickets}
+    -------------------------------------
+
+    ------ Client Query ------
+    {client_query}
+    --------------------------
+
+    Agent :
+    """
+
+    ZENDESK_TEMPLATE_PROMPT = ChatPromptTemplate.from_messages(
+        [
+            SystemMessagePromptTemplate.from_template(system_message_zendesk_template),
+        ]
+    )
+    custom_prompts["ZENDESK_TEMPLATE_PROMPT"] = ZENDESK_TEMPLATE_PROMPT
+
     return custom_prompts
 
 

core/quivr_core/rag/quivr_rag_langgraph.py

@@ -1,5 +1,6 @@
 import asyncio
 import logging
+from collections import OrderedDict
 from typing import (
     Annotated,
     Any,
@@ -13,6 +14,7 @@
     TypedDict,
 )
 from uuid import UUID, uuid4
+
 import openai
 from langchain.retrievers import ContextualCompressionRetriever
 from langchain_cohere import CohereRerank
@@ -752,6 +754,73 @@ def _sort_docs_by_relevance(self, docs: List[Document]) -> List[Document]:
             reverse=True,
         )
 
+    async def retrieve_full_documents_context(self, state: AgentState) -> AgentState:
+        if "tasks" in state:
+            tasks = state["tasks"]
+        else:
+            tasks = UserTasks([state["messages"][0].content])
+
+        if not tasks.has_tasks():
+            return {**state}
+
+        docs = tasks.docs if tasks else []
+
+        relevant_knowledge = {}
+        for doc in docs:
+            knowledge_id = doc.metadata["knowledge_id"]
+            similarity_score = doc.metadata.get("similarity", 0)
+            if knowledge_id in relevant_knowledge:
+                relevant_knowledge[knowledge_id]["count"] += 1
+                relevant_knowledge[knowledge_id]["max_similarity_score"] = max(
+                    relevant_knowledge[knowledge_id]["max_similarity_score"],
+                    similarity_score,
+                )
+                relevant_knowledge[knowledge_id]["chunk_index"] = max(
+                    doc.metadata["chunk_index"],
+                    relevant_knowledge[knowledge_id]["chunk_index"],
+                )
+            else:
+                relevant_knowledge[knowledge_id] = {
+                    "count": 1,
+                    "max_similarity_score": similarity_score,
+                    "chunk_index": doc.metadata["chunk_index"],
+                }
+
+        top_n = min(3, len(relevant_knowledge))
+        # FIXME: Tweak this to return the most relevant knowledges
+        top_knowledge_ids = OrderedDict(
+            sorted(
+                relevant_knowledge.items(),
+                key=lambda x: (
+                    x[1]["max_similarity_score"],
+                    x[1]["count"],
+                ),
+                reverse=True,
+            )[:top_n]
+        )
+
+        logger.info(f"Top knowledge IDs: {top_knowledge_ids}")
+
+        _docs = []
+
+        assert hasattr(
+            self.vector_store, "get_vectors_by_knowledge_id"
+        ), "Vector store must have method 'get_vectors_by_knowledge_id', this is an enterprise only feature"
+
+        for knowledge_id in top_knowledge_ids:
+            _docs.append(
+                await self.vector_store.get_vectors_by_knowledge_id(  # type: ignore
+                    knowledge_id,
+                    end_index=relevant_knowledge[knowledge_id]["chunk_index"],
+                )
+            )
+
+        tasks.set_docs(
+            id=tasks.ids[0], docs=_docs
+        )  # FIXME If multiple IDs is not handled.
+
+        return {**state, "tasks": tasks}
+
     def get_rag_context_length(self, state: AgentState, docs: List[Document]) -> int:
         final_inputs = self._build_rag_prompt_inputs(state, docs)
         msg = custom_prompts.RAG_ANSWER_PROMPT.format(**final_inputs)
@@ -836,6 +905,22 @@ def bind_tools_to_llm(self, node_name: str):
                 return self.llm_endpoint._llm.bind_tools(tools, tool_choice="any")
         return self.llm_endpoint._llm
 
+    def generate_zendesk_rag(self, state: AgentState) -> AgentState:
+        tasks = state["tasks"]
+        docs = tasks.docs if tasks else []
+        messages = state["messages"]
+        user_task = messages[0].content
+        inputs = {
+            "similar_tickets": docs,
+            "client_query": user_task,
+        }
+
+        msg = custom_prompts.ZENDESK_TEMPLATE_PROMPT.format(**inputs)
+        llm = self.bind_tools_to_llm(self.generate_zendesk_rag.__name__)
+        response = llm.invoke(msg)
+
+        return {**state, "messages": [response]}
+
     def generate_rag(self, state: AgentState) -> AgentState:
         tasks = state["tasks"]
         docs = tasks.docs if tasks else []