fix: full fresh acra.py with Jina web search, no duckduckgo

acra.py

@@ -1,9 +1,8 @@
-import os
+import os, httpx
 from google import genai
 from google.genai import types
 from db import supabase
 from classifier_inference import classify_query
-from duckduckgo_search import DDGS
 from typing import List
 
 client      = genai.Client(api_key=os.environ["GEMINI_API_KEY"])
@@ -34,29 +33,35 @@ def adaptive_chunk(text, max_tok=512):
     return chunks or [text]
 
 def web_search(query: str, max_results: int = 5) -> List[dict]:
-    """Web search via Jina AI — no API key, works from cloud, LLM-ready output."""
-    import httpx
+    """Web search via Jina AI s.jina.ai — no API key needed, works from cloud."""
     try:
-        url = f"https://s.jina.ai/?q={httpx.URL(query)}"
-        r   = httpx.get(url,
-                        headers={"Accept": "application/json",
-                                 "X-Respond-With": "no-content"},
-                        timeout=15.0)
+        import urllib.parse
+        encoded = urllib.parse.quote(query)
+        r = httpx.get(
+            f"https://s.jina.ai/?q={encoded}",
+            headers={"Accept": "application/json", "X-Respond-With": "no-content"},
+            timeout=20.0,
+            follow_redirects=True
+        )
         if r.status_code != 200:
-            print(f"Jina search returned {r.status_code}")
+            print(f"Jina returned {r.status_code}: {r.text[:200]}")
             return []
-        data    = r.json()
-        results = data.get("data", [])
-        return [{"title":   item.get("title", ""),
-                 "snippet": item.get("description", item.get("content",""))[:500],
-                 "url":     item.get("url", "")}
-                for item in results[:max_results]
-                if item.get("description") or item.get("content")]
+        data = r.json()
+        items = data.get("data", [])
+        results = []
+        for item in items[:max_results]:
+            snippet = item.get("description") or item.get("content","")
+            if snippet:
+                results.append({
+                    "title":   item.get("title", ""),
+                    "snippet": snippet[:500],
+                    "url":     item.get("url", "")
+                })
+        return results
     except Exception as e:
         print(f"Web search error: {e}")
         return []
 
-
 def decompose(query):
     r = client.models.generate_content(model=GEN_MODEL,
         contents=f"Decompose into 2-4 simpler sub-queries. Numbered list only.\n\nQuery: {query}")
@@ -68,9 +73,8 @@ def compress(query, chunks):
     numbered = "\n\n".join(f"[{i+1}]\n{c}" for i, c in enumerate(chunks))
     r = client.models.generate_content(model=GEN_MODEL, contents=(
         f"You have {len(chunks)} text chunks and a query.\n"
-        f"For each chunk, extract ONLY sentences directly relevant to the query.\n"
-        f"Reply in this exact format:\n"
-        f"[1] <extracted sentences or EMPTY>\n[2] <...> etc.\n\n"
+        f"For each chunk extract ONLY sentences relevant to the query.\n"
+        f"Reply as [1] <text or EMPTY> [2] <text or EMPTY> etc.\n\n"
         f"Query: {query}\n\nChunks:\n{numbered}"))
     import re
     out = []
@@ -88,14 +92,14 @@ def vsearch(query, namespace, user_id, k):
     }).execute().data or [])
 
 PROMPTS = {
-    0: "Answer this from your knowledge:\n\n{q}",
+    0: "Answer from your knowledge:\n\n{q}",
     1: "Answer using ONLY the context. Be concise.\n\nContext:\n{ctx}\n\nQuestion: {q}\nAnswer:",
     2: "Synthesize the context step by step.\n\nContext:\n{ctx}\n\nQuestion: {q}\nAnswer:",
     3: "Use chain-of-thought reasoning.\n\nContext:\n{ctx}\n\nQuestion: {q}\nAnswer:",
 }
 WEB_PROMPT = (
     "Answer the question using ONLY the web search results below.\n"
-    "Be factual and concise. Cite sources where relevant.\n\n"
+    "Be factual and concise.\n\n"
     "Web results:\n{ctx}\n\nQuestion: {q}\nAnswer:"
 )
 
@@ -117,42 +121,42 @@ async def query_pipeline(query, namespace, top_k, rerank, user_id, use_web=False
     level = cls["level"]
     k     = DEPTH[level]
 
-    # ── Web search mode — skip ALL local retrieval, pure web only ──
+    # use_web=True: skip ALL local retrieval, pure Jina web search
     if use_web:
-        web_hits = web_search(query, max_results=6)
-        if not web_hits:
-            return {"answer": "No web results found for this query.",
-                    "sources": [], "complexity": cls, "retrieval_source": "none"}
+        hits = web_search(query, max_results=6)
+        if not hits:
+            return {"answer": "No web results found.", "sources": [],
+                    "complexity": cls, "retrieval_source": "none"}
         ctx = "\n\n---\n\n".join(
-            f"Source: {h['title']}\n{h['snippet']}" for h in web_hits)
+            f"Source: {h['title']}\nURL: {h['url']}\n{h['snippet']}" for h in hits)
         r = client.models.generate_content(model=GEN_MODEL,
             contents=WEB_PROMPT.format(ctx=ctx, q=query))
         return {
             "answer": r.text.strip(),
             "sources": [{"content": h["snippet"][:200],
                          "metadata": {"title": h["title"], "url": h["url"]},
-                         "score": 1.0, "source": "web"} for h in web_hits],
+                         "score": 1.0, "source": "web"} for h in hits],
             "complexity": cls,
             "retrieval_source": "web",
         }
 
-    # ── L0: check docs first, fall back to model knowledge ────────
+    # L0: check docs first, fall back to model knowledge
     if level == 0:
-        hits = vsearch(query, namespace, user_id, 2)
-        if hits:
-            ctx = "\n\n---\n\n".join(h["content"] for h in hits)
+        doc_hits = vsearch(query, namespace, user_id, 2)
+        if doc_hits:
+            ctx = "\n\n---\n\n".join(h["content"] for h in doc_hits)
             r   = client.models.generate_content(model=GEN_MODEL, contents=(
                 f"Use the context if it contains a relevant answer. "
                 f"Otherwise answer from your own knowledge.\n\n"
                 f"Context:\n{ctx}\n\nQuestion: {query}\nAnswer:"))
-            top_score = hits[0].get("similarity", 0)
+            top_score = doc_hits[0].get("similarity", 0)
             return {
                 "answer": r.text.strip(),
                 "sources": [{"content": h["content"][:200],
                              "metadata": h.get("metadata", {}),
                              "score": h.get("similarity", 0),
                              "source": "local"}
-                            for h in hits if h.get("similarity", 0) > 0.5],
+                            for h in doc_hits if h.get("similarity", 0) > 0.5],
                 "complexity": cls,
                 "retrieval_source": "local" if top_score > 0.5 else "model_knowledge",
             }
@@ -161,7 +165,7 @@ async def query_pipeline(query, namespace, top_k, rerank, user_id, use_web=False
         return {"answer": r.text.strip(), "sources": [],
                 "complexity": cls, "retrieval_source": "model_knowledge"}
 
-    # ── L1–L3: local vector retrieval ─────────────────────────────
+    # L1-L3: local vector retrieval
     hits = []
     if level == 3:
         seen = set()
@@ -172,7 +176,6 @@ async def query_pipeline(query, namespace, top_k, rerank, user_id, use_web=False
         hits = vsearch(query, namespace, user_id, k)
 
     if not hits:
-        # Auto-fallback to web if no local docs found
         web_hits = web_search(query, max_results=k)
         if not web_hits:
             return {"answer": "Nothing found locally or on the web.",
@@ -197,8 +200,7 @@ async def query_pipeline(query, namespace, top_k, rerank, user_id, use_web=False
         "sources": [{"content": h["content"][:200], "metadata": h.get("metadata", {}),
                      "score": h.get("similarity", 0), "source": "local"}
                     for h in hits[:len(lc)]],
-        "complexity": cls,
-        "retrieval_source": "local",
+        "complexity": cls, "retrieval_source": "local",
     }
 
 async def run_acra_pipeline(mode, **kw):