M1+M2: RAG-Pipeline mit verbindlichem Grounding

agent/-Paket: Ingest (601 Layer-2-Eintraege -> 3005 Chunks, FTS5-BM25 + Vektoren-Cache), Hybrid-Retrieval (RRF, Stand-Boost, cross_ref-Erweiterung), Ollama-Client (embed/chat, think-Flag-Fallback, kurzes Connect-Budget), Systemprompt mit Zitierpflicht, Post-Validierung (zitierte IDs gemaess Retrieved-Set, 1x Regenerierung, dann Verweigerung), FastAPI (/ask, /health, /reindex), CLI, Goldset (31 Fragen, IDs gegen kb.json verifiziert, inkl. ATZ-Konfliktfall + 4 Verweigerungsfaelle), Eval-Suite, Test-Chat. 41 Offline-Tests gruen. Baseline BM25-only: Hit-Rate 0,871 / Recall@8 0,855 / MRR 0,476. Hybrid-Messung, Antwortmodus-Eval und Modell-Bake-off (M3) auf dem Host ausstaendig (Ollama aus der Zed-Sandbox nicht erreichbar). MEMORY.md und planung.md Umsetzungsstand aktualisiert.
2026-09-14 16:53:04 +02:00
parent bf8191b013
commit 2cba72aeb0
26 changed files with 2574 additions and 1 deletions
@@ -0,0 +1,69 @@
+<!DOCTYPE html>
+<html lang="de">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>PV RAG Agent — Test-Chat</title>
+<style>
+  :root { color-scheme: light dark; }
+  body { font-family: system-ui, sans-serif; max-width: 780px; margin: 2rem auto; padding: 0 1rem; }
+  h1 { font-size: 1.2rem; }
+  #log { display: flex; flex-direction: column; gap: 0.8rem; margin: 1.5rem 0; }
+  .msg { border: 1px solid rgba(128,128,128,.35); border-radius: 8px; padding: .7rem .9rem; white-space: pre-wrap; }
+  .q { background: rgba(128,128,128,.12); }
+  .src { font-size: .78rem; margin-top: .4rem; display: flex; flex-wrap: wrap; gap: .3rem; }
+  .chip { border-radius: 999px; padding: .1rem .5rem; background: rgba(128,128,128,.15); }
+  .warn { border-color: #c08020; }
+  form { display: flex; gap: .5rem; }
+  input { flex: 1; padding: .5rem; border-radius: 6px; border: 1px solid rgba(128,128,128,.5); }
+  button { padding: .5rem 1rem; border-radius: 6px; cursor: pointer; }
+  .meta { font-size: .75rem; opacity: .7; }
+</style>
+</head>
+<body>
+<h1>PV RAG Agent — Wissensbasis Personalverrechnung</h1>
+<p class="meta">Antworten ausschließlich aus der kuratierten Wissensbasis, mit ID- und Stand-Beleg.</p>
+<div id="log"></div>
+<form id="f">
+  <input id="q" placeholder="Frage zur österreichischen Personalverrechnung …" autocomplete="off" required>
+  <button>Senden</button>
+</form>
+<script>
+const log = document.getElementById("log");
+function add(cls, html) {
+  const d = document.createElement("div");
+  d.className = "msg " + cls;
+  d.innerHTML = html;
+  log.appendChild(d);
+}
+document.getElementById("f").addEventListener("submit", async (e) => {
+  e.preventDefault();
+  const q = document.getElementById("q").value.trim();
+  if (!q) return;
+  add("q", q);
+  document.getElementById("q").value = "";
+  try {
+    const r = await fetch("/ask", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ question: q }),
+    });
+    if (!r.ok) {
+      add("warn", "Fehler " + r.status + ": " + (await r.text()));
+      return;
+    }
+    const d = await r.json();
+    const chips = (d.sources || []).map(s =>
+      `<span class="chip">${s.id} · ${s.stand || ""}</span>`).join("");
+    add(d.refused ? "warn" : "",
+      `${d.answer.replace(/</g, "&lt;")}` +
+      (d.sources && d.sources.length ? `<div class="src">${chips}</div>` : "") +
+      `<div class="meta">Modell: ${d.model} · ${d.latency_ms} ms · ` +
+      `${d.verified ? "zitiergeprüft" : "UNVERIFIZIERT"}${d.refused ? " · verweigert" : ""}</div>`);
+  } catch (err) {
+    add("warn", "Netzwerkfehler: " + err);
+  }
+});
+</script>
+</body>
+</html>