feat: search_debug(), RAG details in web logs

bot.py

@@ -7,6 +7,7 @@ from openai import OpenAI
 import httpx
 from dotenv import load_dotenv
 from log_store import add as log_add
+from knowledge import search_debug
 
 load_dotenv()
 
@@ -15,7 +16,7 @@ logger = logging.getLogger(__name__)
 GROQ_API_KEY = os.getenv('GROQ_API_KEY')
 LLAMA_MODEL_PATH = os.getenv('LLAMA_MODEL_PATH')
 
-from knowledge import reload_knowledge, search
+from knowledge import reload_knowledge
 kb_items, kb_tfidf = reload_knowledge(['knowledge_base.json', 'knowledge_base_insoles.json'])
 
 _llm = None
@@ -145,7 +146,7 @@ def start_bot(bot_token: str):
                         "chat_id": chat_id,
                         "text": msg,
                     })
-                    log_add(text, msg, [], "greeting")
+                    log_add(question=text, response=msg, mode="greeting", search_method="—", timing_ms=0)
                     continue
 
                 if need_operator(text):
@@ -154,26 +155,41 @@ def start_bot(bot_token: str):
                         "chat_id": chat_id,
                         "text": msg,
                     })
-                    log_add(text, msg, [], "operator")
+                    log_add(question=text, response=msg, mode="operator", search_method="—", timing_ms=0)
                     continue
 
-                results = search(text, top_k=2, items=kb_items, tfidf=kb_tfidf)
-                sources = [r.title for r in results]
+                t0 = time.time()
+                debug = search_debug(text, top_k=2)
+                results = debug["items"]
+                t1 = time.time()
 
                 response = get_grok_response(text, results)
                 mode = "groq"
+                llm_model = "llama-3.3-70b-versatile (Groq)"
                 if response is None:
                     response = get_local_response(text, results)
                     mode = "local"
+                    llm_model = "Llama 3.2 3B (local)"
                 if response is None:
                     response = "Извините, не удалось получить ответ. Попробуйте позже или позвоните +375 (29) 145-03-03."
                     mode = "fallback"
+                    llm_model = ""
+                t2 = time.time()
 
                 client.post(f"{api_url}/sendMessage", json={
                     "chat_id": chat_id,
                     "text": response,
                 })
-                log_add(text, response, sources, mode)
+
+                log_add(
+                    question=text,
+                    response=response,
+                    mode=mode,
+                    search_method=debug["method"],
+                    search_details=debug["details"],
+                    llm_model=llm_model,
+                    timing_ms=round((t1 - t0) * 1000 + (t2 - t1) * 1000),
+                )
 
         except Exception as e:
             logger.error(f"Polling error: {e}")

knowledge.py

@@ -128,16 +128,19 @@ def load_knowledge_base(paths: list[str] | str) -> list[KnowledgeItem]:
 
 def search(query: str, top_k: int = 2, items: list[KnowledgeItem] = None,
            tfidf: dict = None) -> list[KnowledgeItem]:
+    return search_debug(query, top_k)["items"]
+
+
+def search_debug(query: str, top_k: int = 2) -> dict:
+    """Returns {items, method, details: [{title, bm25_score, embed_rank, rrf_score}]}"""
     n = len(_sections)
     if n == 0:
-        return []
+        return {"items": [], "method": "none", "details": []}
 
-    # --- BM25 scores ---
     bm25_scores = None
     if _bm25 is not None:
         bm25_scores = _bm25.get_scores(_tokenize(query))
 
-    # --- Dense scores (embedding) ---
     embed_ranks = None
     qvec = _embed_query(query)
     if qvec is not None and _index is not None and _index.ntotal > 0:
@@ -148,33 +151,42 @@ def search(query: str, top_k: int = 2, items: list[KnowledgeItem] = None,
     if bm25_scores is not None and embed_ranks is not None:
         rrf = {}
         for i in range(n):
-            score = 0.0
-            # BM25 rank
             bm25_rank = sorted(range(n), key=lambda j: -bm25_scores[j]).index(i)
-            score += 0.4 * (1 / (bm25_rank + 1))
-            # Embed rank
+            rrf_score = 0.0
+            rrf_score += 0.4 * (1 / (bm25_rank + 1))
             if i in embed_ranks:
-                score += 0.6 * (1 / (embed_ranks[i] + 1))
-            rrf[i] = score
+                rrf_score += 0.6 * (1 / (embed_ranks[i] + 1))
+            rrf[i] = rrf_score
         top_indices = sorted(rrf.keys(), key=lambda i: -rrf[i])[:top_k]
-        return [_sections[i] for i in top_indices]
+        details = [{
+            "id": _sections[i].id,
+            "title": _sections[i].title,
+            "bm25_rank": sorted(range(n), key=lambda j: -bm25_scores[j]).index(i),
+            "embed_rank": embed_ranks.get(i, None),
+            "rrf_score": round(rrf[i], 4),
+        } for i in top_indices]
+        return {"items": [_sections[i] for i in top_indices], "method": "hybrid (bge-m3+BM25)", "details": details}
 
     # --- Embedding only ---
     if embed_ranks is not None:
         top_indices = sorted(embed_ranks.keys(), key=lambda i: embed_ranks[i])[:top_k]
-        return [_sections[i] for i in top_indices]
+        details = [{"id": _sections[i].id, "title": _sections[i].title, "embed_rank": embed_ranks[i], "rrf_score": None} for i in top_indices]
+        return {"items": [_sections[i] for i in top_indices], "method": "bge-m3 only", "details": details}
 
     # --- BM25 only ---
     if bm25_scores is not None:
         top_indices = sorted(range(n), key=lambda i: -bm25_scores[i])[:top_k]
-        return [_sections[i] for i in top_indices]
+        details = [{"id": _sections[i].id, "title": _sections[i].title, "bm25_rank": i, "rrf_score": None} for i in top_indices]
+        return {"items": [_sections[i] for i in top_indices], "method": "BM25 only", "details": details}
 
     # --- TF-IDF fallback ---
     if _tfidf_backup is not None and _items_backup:
         from knowledge_tfidf_backup import search as tfidf_search
-        return tfidf_search(query, top_k=top_k, items=_items_backup, tfidf=_tfidf_backup)
+        items = tfidf_search(query, top_k=top_k, items=_items_backup, tfidf=_tfidf_backup)
+        details = [{"id": it.id, "title": it.title, "rrf_score": None} for it in items]
+        return {"items": items, "method": "TF-IDF fallback", "details": details}
 
-    return []
+    return {"items": [], "method": "none", "details": []}
 
 
 def reload_knowledge(paths: list[str] | str) -> tuple[list[KnowledgeItem], dict]:

log_store.py

@@ -1,19 +1,23 @@
 """Thread-safe in-memory log of bot interactions."""
 
-import threading
+import threading, time
 from collections import deque
 
 _log: deque[dict] = deque(maxlen=50)
 _lock = threading.Lock()
 
 
-def add(question: str, response: str, sources: list[str], mode: str):
+def add(**kwargs):
     entry = {
-        "time": __import__("time").strftime("%H:%M:%S", __import__("time").gmtime()),
-        "question": question,
-        "response": response[:300],
-        "sources": sources,
-        "mode": mode,
+        "time": time.strftime("%H:%M:%S", time.gmtime()),
+        "question": kwargs.get("question", ""),
+        "response": str(kwargs.get("response", ""))[:500],
+        "mode": kwargs.get("mode", ""),
+        "search_method": kwargs.get("search_method", ""),
+        "search_details": kwargs.get("search_details", []),
+        "llm_model": kwargs.get("llm_model", ""),
+        "timing_ms": kwargs.get("timing_ms", 0),
+        "error": kwargs.get("error", ""),
     }
     with _lock:
         _log.appendleft(entry)

server.py

@@ -36,29 +36,59 @@ class HealthHandler(BaseHTTPRequestHandler):
 <meta charset="utf-8"><title>ORTOS Bot Logs</title>
 <style>
 body{font-family:sans-serif;margin:20px;background:#111;color:#eee}
-table{border-collapse:collapse;width:100%}
-th,td{text-align:left;padding:8px;border-bottom:1px solid #333;vertical-align:top}
-th{background:#222;color:#0f0}
+table{border-collapse:collapse;width:100%;font-size:13px}
+th,td{text-align:left;padding:6px 10px;border-bottom:1px solid #333;vertical-align:top}
+th{background:#222;color:#0f0;position:sticky;top:0}
 tr:hover{background:#1a1a1a}
-.mode{font-weight:bold;padding:2px 6px;border-radius:3px;font-size:12px}
+.mode{font-weight:bold;padding:2px 6px;border-radius:3px;font-size:11px;white-space:nowrap}
 .groq{background:#1a3a1a;color:#4f4}
-.local{background:#3a1a1a;color:#f44}
+.local{background:#3a1a1a;color:#f88}
 .fallback{background:#3a3a1a;color:#ff4}
 .greeting{background:#1a1a3a;color:#44f}
 .operator{background:#3a1a3a;color:#f4f}
-.src{color:#888;font-size:11px}
-.q{color:#ffa;max-width:300px;word-break:break-word}
-.r{color:#afa;max-width:400px;word-break:break-word}
+.q{color:#ffa;max-width:250px;word-break:break-word}
+.r{color:#afa;max-width:350px;word-break:break-word}
+.detail{color:#888;font-size:11px;margin-top:4px;border-top:1px solid #333;padding-top:4px}
+.lbl{color:#666}
+.val{color:#eee}
+.src{color:#8af}
+.err{color:#f44}
+summary{cursor:pointer;color:#8af;font-size:12px}
 </style></head><body>
 <h2>ORTOS Bot &mdash; last 50 interactions</h2>
-<table><tr><th>Time</th><th>Mode</th><th>Question</th><th>Response</th><th>Sources</th></tr>"""
+<table><thead><tr><th>Time</th><th>Mode</th><th>Q</th><th>Response</th><th>RAG</th></tr></thead>"""
             for e in entries:
                 css = e["mode"]
-                src = ", ".join(e["sources"]) if e["sources"] else "—"
+                md = e["search_method"]
+                llm = html.escape(e["llm_model"])
+                ms = e["timing_ms"]
+
+                # Search details
+                rag_html = f"<span class='src'>{html.escape(md)}</span>"
+                rag_html += f"<br><span class='lbl'>LLM:</span> <span class='val'>{llm}</span>"
+                rag_html += f"<br><span class='lbl'>⏱</span> <span class='val'>{ms}ms</span>"
+
+                if e.get("search_details"):
+                    rag_html += "<details><summary>search results</summary>"
+                    for d in e["search_details"]:
+                        rag_html += f"<div class='detail'>"
+                        rag_html += f"<b class='src'>{html.escape(d.get('title',''))}</b>"
+                        br = d.get("bm25_rank")
+                        er = d.get("embed_rank")
+                        rs = d.get("rrf_score")
+                        if br is not None:
+                            rag_html += f"<br><span class='lbl'>BM25 rank:</span> <span class='val'>{br}</span>"
+                        if er is not None:
+                            rag_html += f"<br><span class='lbl'>bge-m3 rank:</span> <span class='val'>{er}</span>"
+                        if rs is not None:
+                            rag_html += f"<br><span class='lbl'>RRF score:</span> <span class='val'>{rs}</span>"
+                        rag_html += "</div>"
+                    rag_html += "</details>"
+
                 page += f"<tr><td>{e['time']}</td><td><span class='mode {css}'>{e['mode']}</span></td>"
                 page += f"<td class='q'>{html.escape(e['question'])}</td>"
                 page += f"<td class='r'>{html.escape(e['response'])}</td>"
-                page += f"<td class='src'>{src}</td></tr>"
+                page += f"<td>{rag_html}</td></tr>"
             page += "</table></body></html>"
 
             self.send_response(200)