Include project knowledge in Workflow

api/routes/completion.py

@@ -116,7 +116,8 @@ async def _complete(request: CompletionRequest) -> CompletionResponse:
         agent_type=request.agent_type,
         temperature=request.temperature,
         max_tokens=request.max_tokens,
-        conversation_history=request.conversation_history
+        conversation_history=request.conversation_history,
+        project_id=request.project_id
     )
     
     return CompletionResponse(**result)
@@ -141,7 +142,8 @@ async def _stream_completion(request: CompletionRequest) -> StreamingResponse:
                 agent_type=request.agent_type,
                 temperature=request.temperature,
                 max_tokens=request.max_tokens,
-                conversation_history=request.conversation_history
+                conversation_history=request.conversation_history,
+                project_id=request.project_id,
             ):
                 # Format as SSE: "data: {json}\n\n"
                 chunk_json = json.dumps(chunk, ensure_ascii=False)

domain/models.py

@@ -35,6 +35,8 @@ class CompletionRequest(BaseModel):
         default=None,
         description="Optional conversation history"
     )
+    # Project-scoped retrieval
+    project_id: Optional[str] = Field(default=None, description="Optional project id to scope retrieval")
 
 
 class CompletionResponse(BaseModel):

graphs/agents/chat_agent.py

@@ -17,6 +17,19 @@ def chat_node(llm: BaseChatModel) -> Callable[[AgentState], AgentState]:
 
         formation_context = state.get("formation_context", "")
         prestation_context = state.get("prestation_context", "")
+        project_context = state.get("project_context", "")
+
+        
+        if project_context:
+            sys_msgs.append(
+                SystemMessage(
+                    content=(
+                        "CONTEXTE PROJET (extraits des documents du projet; n'utilise rien d'autre):\n\n"
+                        f"{project_context}\n\n"
+                        "Consignes projet: Ce contenu indique des informations complémentaires à prendre en compte pour répondre à la question. "
+                    )
+                )
+            )
 
         if formation_context:
             sys_msgs.append(

graphs/base_graph.py

@@ -23,6 +23,8 @@ class AgentState(TypedDict, total=False):
     prestation_docs: List[Document]
     formation_context: str
     prestation_context: str
+    project_docs: List[Document]
+    project_context: str
 
 
 def create_simple_graph(llm: BaseChatModel):

graphs/nodes/retrieval.py

@@ -1,11 +1,12 @@
 """Retrieval nodes for LangGraph workflows."""
-from typing import Dict
+from typing import Dict, Optional
 
 from graphs.state import AgentState
-from retrievers.supabase import get_retriever, format_documents
+from retrievers.supabase import get_retriever, get_retriever_for, format_documents
+from langchain_core.messages import SystemMessage
 
 
-def retrieve_both_types(state: AgentState) -> AgentState:
+def retrieve_catalogue(state: AgentState) -> AgentState:
     """Builds a query from the state and retrieves formation and prestation docs.
 
     Returns the augmented state with docs and formatted contexts.
@@ -18,14 +19,29 @@ def retrieve_both_types(state: AgentState) -> AgentState:
                 query_text = (msg.content or "").strip()
                 break
 
+    # retrievers
     formation_retriever = get_retriever("formation", k=8)
     prestation_retriever = get_retriever("prestation", k=8)
 
+    # Invoke
     formation_docs = formation_retriever.invoke(query_text)
     prestation_docs = prestation_retriever.invoke(query_text)
 
-    formation_context = format_documents(formation_docs, "formation")
-    prestation_context = format_documents(prestation_docs, "prestation")
+    new_formation_context = format_documents(formation_docs, "formation")
+    new_prestation_context = format_documents(prestation_docs, "prestation")
+
+    # Merge with any existing contexts (e.g., from project retrieval) to keep both
+    old_formation_context = state.get("formation_context", "")
+    old_prestation_context = state.get("prestation_context", "")
+
+    formation_context = (
+        (old_formation_context + "\n\n---\n\n" + new_formation_context).strip()
+        if old_formation_context else new_formation_context
+    )
+    prestation_context = (
+        (old_prestation_context + "\n\n---\n\n" + new_prestation_context).strip()
+        if old_prestation_context else new_prestation_context
+    )
 
     return {
         "formation_docs": formation_docs,
@@ -35,3 +51,38 @@ def retrieve_both_types(state: AgentState) -> AgentState:
     }
 
 
+def retrieve_projects(state: AgentState) -> AgentState:
+    """Retrieve only project-scoped documents (formation and prestation) and add a system hint.
+
+    Used when `project_id` is present to focus retrieval on the 'projects' vector index.
+    """
+    # Extract user query
+    query_text = state.get("query") or ""
+    if not query_text:
+        for msg in reversed(list(state.get("messages", []))):
+            if getattr(msg, "type", "") == "human":
+                query_text = (msg.content or "").strip()
+                break
+
+    project_id: Optional[str] = state.get("project_id")  # type: ignore[assignment]
+    index_name: Optional[str] = "projects"  # type: ignore[assignment]
+
+    # Safety: if no project_id, return state unchanged (router should avoid calling us)
+    if not project_id:
+        return {}
+
+    extra = {"project_id": project_id}
+    
+    project_retriever = get_retriever_for(index_name, k=8, filter=extra)
+
+    project_docs = project_retriever.invoke(query_text)
+
+    projet_context = format_documents(project_docs, "project")
+
+
+    return {
+        "project_docs": project_docs,
+        "project_context": projet_context,
+    }
+
+

graphs/state.py

@@ -14,13 +14,17 @@ class AgentState(TypedDict, total=False):
     # Conversation
     messages: Annotated[Sequence[BaseMessage], add_messages]
     query: Optional[str]
+    # Project scoping
+    project_id: Optional[str]
+    index_name: Optional[str]
 
     # RAG retrieval results
     formation_docs: List[Document]
     prestation_docs: List[Document]
     formation_context: str
     prestation_context: str
-
+    project_docs: List[Document]
+    project_context: str
     # Summarization artifacts
     summary_markdown: str
     summary_pdf_path: str  # local path or URL if uploaded

graphs/workflows/orchestrated.py

@@ -4,7 +4,7 @@ from langchain_core.language_models.chat_models import BaseChatModel
 
 from graphs.state import AgentState
 from graphs.agents.classifier_agent import classifier_node
-from graphs.nodes.retrieval import retrieve_both_types
+from graphs.nodes.retrieval import retrieve_catalogue, retrieve_projects
 from graphs.agents.chat_agent import chat_node
 from graphs.agents.summarizer_agent import summarizer_llm_node, summarizer_export_node
 # from tools.markdown import markdown_to_html
@@ -17,7 +17,14 @@ def create_orchestrated_graph(llm: BaseChatModel):
 
     # Nodes
     workflow.add_node("classify", classifier_node(llm))
-    workflow.add_node("retrieve", retrieve_both_types)
+    workflow.add_node("retrieve", retrieve_catalogue)
+    # Route to classic vs project retrieval
+    def _router_passthrough(state: AgentState) -> AgentState:
+        # Must write at least one allowed key; pass through the current query
+        q = state.get("query") or ""
+        return {"query": q}
+    workflow.add_node("retrieve_router", _router_passthrough)
+    workflow.add_node("retrieve_project", retrieve_projects)
     workflow.add_node("agent", chat_node(llm))
     workflow.add_node("summarizer_llm", summarizer_llm_node(llm))
     workflow.add_node(
@@ -37,15 +44,29 @@ def create_orchestrated_graph(llm: BaseChatModel):
         "classify",
         lambda s: getattr(s.get("classification"), "classification", "CLASSIC"),
         {
-            "CLASSIC": "retrieve",
+            # Route through a retrieval router to optionally branch to project retrieval
+            "CLASSIC": "retrieve_router",
             "SUMMARIZE": "summarizer_llm",
-            "UNKNOWN": "retrieve",
+            "UNKNOWN": "retrieve_router",
+        },
+    )
+
+    # Conditional choice between project vs classic retrieval
+    workflow.add_conditional_edges(
+        "retrieve_router",
+        lambda s: "PROJECT" if s.get("project_id") else "CLASSIC",
+        {
+            "PROJECT": "retrieve_project",
+            "CLASSIC": "retrieve",
         },
     )
 
     # Linear branches
+    # If project path is taken, run project retrieval then classic retrieval
+    workflow.add_edge("retrieve_project", "retrieve")
     workflow.add_edge("retrieve", "agent")
     workflow.add_edge("agent", END)
+    
     workflow.add_edge("summarizer_llm", "summarizer_export")
     workflow.add_edge("summarizer_export", END)
 

retrievers/supabase.py

@@ -5,6 +5,7 @@ import os
 
 from langchain_core.documents import Document
 from langchain_mistralai import MistralAIEmbeddings
+from langchain_openai import OpenAIEmbeddings
 from langchain_community.vectorstores import SupabaseVectorStore
 from supabase import create_client, Client
 
@@ -31,7 +32,7 @@ def get_retriever(doc_type: str, k: Optional[int] = None):
 
     client: Client = create_client(url, key)
     vector_store = SupabaseVectorStore(
-        embedding=MistralAIEmbeddings(model="mistral-embed", api_key=settings.mistralai_api_key),
+        embedding=OpenAIEmbeddings(),
         client=client,
         table_name=settings.supabase_table,
         query_name=settings.supabase_match_fn,
@@ -42,12 +43,17 @@ def get_retriever(doc_type: str, k: Optional[int] = None):
     )
 
 
-def get_retriever_for(index_name: str, doc_type: str, k: Optional[int] = None):
+def get_retriever_for(
+    index_name: str,
+    k: Optional[int] = None,
+    filter: Optional[dict] = None,
+):
     """Return a retriever for a specific logical index (table/query pair)."""
     vector_store = get_vector_store(index_name=index_name)
     top_k = int(k or settings.rag_top_k)
+    
     return vector_store.as_retriever(
-        search_kwargs={"k": top_k, "filter": {"type": doc_type}}
+        search_kwargs={"k": top_k, "filter": filter}
     )
 
 

services/agent_service.py

@@ -32,7 +32,8 @@ class AgentService:
         agent_type: AgentType = AgentType.SIMPLE,
         temperature: float = 0.7,
         max_tokens: Optional[int] = None,
-        conversation_history: Optional[List[Dict[str, str]]] = None
+        conversation_history: Optional[List[Dict[str, str]]] = None,
+        project_id: Optional[str] = None
     ) -> dict:
         """
         Invoke agent for a single response (non-streaming).
@@ -65,7 +66,11 @@ class AgentService:
         
         # Execute graph with latency
         start_time = time.time()
-        result = await graph.ainvoke({"messages": messages})
+        result = await graph.ainvoke({
+            "messages": messages,
+            "query": message,
+            "project_id": project_id
+        })
         latency_s = time.time() - start_time
         
         # Extract response
@@ -106,7 +111,8 @@ class AgentService:
         agent_type: AgentType = AgentType.SIMPLE,
         temperature: float = 0.7,
         max_tokens: Optional[int] = None,
-        conversation_history: Optional[List[Dict[str, str]]] = None
+        conversation_history: Optional[List[Dict[str, str]]] = None,
+        project_id: Optional[str] = None
     ) -> AsyncIterator[dict]:
         """
         Stream agent response token by token.
@@ -144,7 +150,11 @@ class AgentService:
         documents = []
 
         # Stream graph execution
-        async for msg in graph.astream({"messages": messages}, stream_mode=["messages","updates"]):
+        async for msg in graph.astream({
+            "messages": messages,
+            "query": message,
+            "project_id": project_id
+        }, stream_mode=["messages","updates"]):
             # LangGraph may yield (node_name, message) tuples in messages mode
             event = None
             params = None