retrival einbinden

math-tutor/backend/app/api/chat.py

@@ -3,11 +3,14 @@ from __future__ import annotations
 from typing import List, Optional
 
 import os
+import httpx
 from fastapi import APIRouter, HTTPException
 from dotenv import load_dotenv
-import httpx
 from pydantic import BaseModel, Field
 
+from app.services.embeddings import OpenAILikeEmbeddings
+from app.services import vector_store
+
 router = APIRouter()
 
 load_dotenv()
@@ -17,16 +20,19 @@ OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
 OPENAI_MODEL = os.getenv("OPENAI_MODEL", "gpt-4o-mini")
 
 SYSTEM_PROMPT = """Du bist ein Mathe-Tutor. Antworte auf Deutsch, klar und korrekt.
-Nutze ausschließlich den bereitgestellten Kontext.
-Gib wenn möglich eine kurze Struktur: 
-(1) Idee, 
-(2) Definition, 
-(3) kurzer Begründungs-/Rechenweg, 
-(4) Mini-Beispiel.
+Nutze ausschließlich den bereitgestellten Kontext. Wenn nichts zur Frage im Kontext steht, antworte mit "Dazu steht nichts im Material" und nichts weiter.
+Gib wenn möglich eine kurze Struktur: (1) Idee, (2) Definition, (3) kurzer Begründungs-/Rechenweg, (4) Mini-Beispiel.
 Zitiere Quellen inline mit den eckigen Klammern, die im Kontext vorangestellt sind, z.B. [s2/ss1/c3 | definition | ...].
-Wenn nichts zur Frage im Kontext steht, antworte mit "Dazu steht nichts im Material" und gib nichts weiter aus.
 """
 
+CONTEXT_LIMITS = {
+    "direct": 4,
+    "indirect": 6,
+    "subsection": 2,
+    "section": 1,
+}
+
+
 class ChatMessage(BaseModel):
     role: str = Field(..., pattern="^(user|assistant)$")
     text: str = Field(..., min_length=1)
@@ -42,6 +48,59 @@ class ChatResponse(BaseModel):
     sources: List[str] = []
 
 
+def _get_embedder() -> OpenAILikeEmbeddings:
+    base_url = os.getenv("OPENAI_BASE_URL")
+    api_key = os.getenv("OPENAI_API_KEY")
+    model = os.getenv("OPENAI_EMBED_MODEL", "text-embedding-3-large")
+    if not base_url or not api_key:
+        raise HTTPException(status_code=500, detail="Missing OPENAI_BASE_URL or OPENAI_API_KEY")
+    return OpenAILikeEmbeddings(base_url=base_url, api_key=api_key, model=model, target_dim=1024)
+
+
+def _format_ref(doc: vector_store.Retrieved) -> str:
+    meta = doc.metadata
+    sec = meta.get("section_index")
+    sub = meta.get("subsection_index")
+    child = meta.get("child_index")
+    ref = []
+    if sec is not None:
+        ref.append(f"s{sec}")
+    if sub is not None:
+        ref.append(f"ss{sub}")
+    if child is not None:
+        ref.append(f"c{child}")
+    ref_id = "/".join(ref) if ref else "unknown"
+
+    doc_type = meta.get("type") or doc.doc_type
+    title = (
+        meta.get("title")
+        or meta.get("subsection_title")
+        or meta.get("section_title")
+        or meta.get("path")
+        or "Untitled"
+    )
+    return f"[{ref_id} | {doc_type} | {title}]"
+
+
+def _build_context(result: dict) -> str:
+    blocks: List[str] = []
+
+    def add_group(label: str, items: List[vector_store.Retrieved], limit: int) -> None:
+        if not items:
+            return
+        for doc in items[:limit]:
+            blocks.append(f"{label} {_format_ref(doc)}\n{doc.markdown}")
+
+    add_group("DIRECT", result.get("children_direct", []), CONTEXT_LIMITS["direct"])
+    add_group("INDIRECT", result.get("children_expanded", []), CONTEXT_LIMITS["indirect"])
+    add_group("SUBSECTION", result.get("subsections", []), CONTEXT_LIMITS["subsection"])
+    add_group("SECTION", result.get("sections", []), CONTEXT_LIMITS["section"])
+
+    if not blocks:
+        return "KONTEXT: (leer)"
+    return "KONTEXT:\n" + "\n\n".join(blocks)
+
+
 @router.post("/api/chat", response_model=ChatResponse)
 def chat(request: ChatRequest) -> ChatResponse:
     if not request.messages:
@@ -52,10 +111,25 @@ def chat(request: ChatRequest) -> ChatResponse:
         reply = f"Mock reply to: {last.text}"
         return ChatResponse(reply=reply, sources=["doc:example"])
 
-    messages_payload = [{"role": "system", "content": SYSTEM_PROMPT}] + [
-        {"role": message.role, "content": message.text}
-        for message in request.messages
-    ]
+    pg_url = os.getenv("POSTGRES_URL")
+    if not pg_url:
+        raise HTTPException(status_code=500, detail="Missing POSTGRES_URL")
+
+    query_text = request.messages[-1].text
+    embedder = _get_embedder()
+    retrieval = vector_store.retrieve(
+        pg_url=pg_url,
+        embedder=embedder,
+        query=query_text,
+        k=8,
+        expand_links=True,
+    )
+    context_block = _build_context(retrieval)
+
+    messages_payload = [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "system", "content": context_block},
+    ] + [{"role": message.role, "content": message.text} for message in request.messages]
 
     payload = {
         "model": OPENAI_MODEL,