"""
KG Embedding Server — Unified Smart Search
FastAPI on HuggingFace Spaces
Flow at startup:
  1. Load embedding model (sentence-transformers) + FAISS + numpy
  2. Scan KNOWLEDGE_PATH -> chunk -> embed -> build FAISS index
  3. Initialise web-search provider (Tavily / Jina)
  4. Initialise LLM (Cerebras / Cohere) via LangChain
Endpoints:
  POST /search   — UNIFIED INTELLIGENT SEARCH (Local KB + Web + LLM/Agent)
  POST /reload   — re-scan folder and rebuild FAISS index
  GET  /health   — status check (includes web/LLM status)
  GET  /debug    — list knowledge base files
  GET  /ping     — keep-alive
"""

from fastapi import FastAPI, HTTPException, Header, Depends
from pydantic import BaseModel
from typing import Any, Optional
import os
import re
import json
import contextlib
import urllib.request
import urllib.parse
import hmac  # 🔒 لمقارنة الـ API Key بشكل آمن
from io import StringIO

app = FastAPI(title="KG Embedding Server — Unified Smart Search")

# ══════════════════════════════════════════════════════
#  CONFIG
# ══════════════════════════════════════════════════════
# Local KB
KNOWLEDGE_PATH = "./Groovy Tutorials"
CHUNK_SIZE_WORDS = int(os.environ.get("CHUNK_SIZE_WORDS", 300))
CHUNK_OVERLAP_WORDS = int(os.environ.get("CHUNK_OVERLAP_WORDS", 50))
EMBED_MODEL_NAME = os.environ.get("EMBED_MODEL_NAME", "all-MiniLM-L6-v2")
ALLOWED_EXTENSIONS = (".md", ".txt")

# Web Search
TAVILY_API_KEY = os.environ.get("TAVILY_API_KEY", "")
JINA_API_KEY = os.environ.get("JINA_API_KEY", "")
WEB_SEARCH_PROVIDER = os.environ.get("WEB_SEARCH_PROVIDER", "tavily")
MAX_WEB_RESULTS = int(os.environ.get("MAX_WEB_RESULTS", 3))
MAX_WEB_CONTENT_CHARS = int(os.environ.get("MAX_WEB_CONTENT_CHARS", 1500))
WEB_INCLUDE_DOMAINS = [
    d.strip() for d in os.environ.get(
        "WEB_INCLUDE_DOMAINS",
        "community.atlassian.com,support.atlassian.com,docs.atlassian.com,"
        "developer.atlassian.com,adaptavist.com,scriptrunner-docs.adaptavist.com,"
        "blog.adaptavist.com,jira.atlassian.com"
    ).split(",") if d.strip()
]

# LLM
CEREBRAS_API_KEY = os.environ.get("CEREBRAS_API_KEY", "")
COHERE_API_KEY = os.environ.get("COHERE_API_KEY", "")
LLM_PROVIDER = os.environ.get("LLM_PROVIDER", "cerebras")
LLM_MODEL = os.environ.get("LLM_MODEL", "")
LLM_TEMPERATURE = float(os.environ.get("LLM_TEMPERATURE", 0.1))
MAX_FINAL_ANSWER_TOKENS = int(os.environ.get("MAX_FINAL_ANSWER_TOKENS", 2000))
if not LLM_MODEL:
    LLM_MODEL = "llama3.1-70b" if LLM_PROVIDER == "cerebras" else "command-r-plus-08-2024"

# Agent
USE_AGENT = os.environ.get("USE_AGENT", "false").lower() == "true"
AGENT_MAX_ITERATIONS = int(os.environ.get("AGENT_MAX_ITERATIONS", 4))

# ── Auth (API Key) ──
API_SECRET_KEY = os.environ.get("API_SECRET_KEY", "")
AUTH_ENABLED = bool(API_SECRET_KEY)

# ══════════════════════════════════════════════════════
#  GLOBALS
# ══════════════════════════════════════════════════════
_model: Any = None
_use_st = False
_faiss: Any = None
_np: Any = None
_index: Any = None
_chunks: list[dict] = []

_llm: Any = None
_llm_provider_name: Optional[str] = None
_web_search_available = False
_web_search_provider_name: Optional[str] = None
_agent_executor: Any = None

@app.on_event("startup")
def startup():
    global _model, _use_st, _faiss, _np

    if AUTH_ENABLED:
        print("[Server] API Key Authentication: ENABLED ✓")
    else:
        print("[Server] API Key Authentication: DISABLED (Running in open mode)")

    try:
        from sentence_transformers import SentenceTransformer
        with contextlib.redirect_stdout(StringIO()), contextlib.redirect_stderr(StringIO()):
            _model = SentenceTransformer(EMBED_MODEL_NAME)
        _use_st = True
        print(f"[Server] embedding model '{EMBED_MODEL_NAME}' loaded ✓")
    except Exception as e:
        print(f"[Server] sentence-transformers unavailable: {e}")

    try:
        import faiss as _faiss_mod
        _faiss = _faiss_mod
        print("[Server] faiss loaded ✓")
    except Exception as e:
        print(f"[Server] faiss unavailable: {e}")

    try:
        import numpy as np
        _np = np
        print("[Server] numpy loaded ✓")
    except Exception as e:
        print(f"[Server] numpy unavailable: {e}")

    build_index()
    init_web_search()
    init_llm()

    if USE_AGENT and _llm is not None:
        try:
            _create_agent_executor()
        except Exception as e:
            print(f"[Server] agent creation failed: {e}")


# ══════════════════════════════════════════════════════
#  AUTH — API Key verification
# ══════════════════════════════════════════════════════
def verify_api_key(
    x_api_key: Optional[str] = Header(None, alias="X-API-Key"),
    authorization: Optional[str] = Header(None)
):
    """
    بيقبل المفتاح إما:
      - Header:  X-API-Key: <SECRET>
      - Header:  Authorization: Bearer <SECRET>
    لو API_SECRET_KEY فاضي على السيرفر → بيمشى عادي (graceful degradation).
    """
    if not API_SECRET_KEY:
        return  # السيرفر مش متظبط عليه مفتاح → اسمح بكل الطلبات

    provided = x_api_key
    if not provided and authorization:
        parts = authorization.split(" ", 1)
        if len(parts) == 2 and parts[0].lower() == "bearer":
            provided = parts[1].strip()

    if not provided or not hmac.compare_digest(provided, API_SECRET_KEY):
        raise HTTPException(
            status_code=401,
            detail="Invalid or missing API key",
            headers={"WWW-Authenticate": 'ApiKey realm="kg-embedding"'},
        )


# ══════════════════════════════════════════════════════
#  INIT FUNCTIONS
# ══════════════════════════════════════════════════════
def init_web_search():
    global _web_search_available, _web_search_provider_name
    if TAVILY_API_KEY:
        _web_search_available = True
        _web_search_provider_name = "tavily"
        print("[Server] web search: Tavily ✓")
    elif JINA_API_KEY:
        _web_search_available = True
        _web_search_provider_name = "jina"
        print("[Server] web search: Jina ✓")
    else:
        print("[Server] web search: not configured")

def init_llm():
    global _llm, _llm_provider_name
    try:
        if LLM_PROVIDER == "cerebras" and CEREBRAS_API_KEY:
            from langchain_cerebras import ChatCerebras
            _llm = ChatCerebras(
                model=LLM_MODEL,
                api_key=CEREBRAS_API_KEY,
                temperature=LLM_TEMPERATURE,
                max_tokens=MAX_FINAL_ANSWER_TOKENS,
            )
            _llm_provider_name = "cerebras"
            print(f"[Server] LLM loaded: Cerebras ({LLM_MODEL}) ✓")
        elif LLM_PROVIDER == "cohere" and COHERE_API_KEY:
            from langchain_cohere import ChatCohere
            _llm = ChatCohere(
                model=LLM_MODEL,
                cohere_api_key=COHERE_API_KEY,
                temperature=LLM_TEMPERATURE,
                max_tokens=MAX_FINAL_ANSWER_TOKENS,
            )
            _llm_provider_name = "cohere"
            print(f"[Server] LLM loaded: Cohere ({LLM_MODEL}) ✓")
        else:
            print("[Server] LLM not configured.")
    except Exception as e:
        print(f"[Server] LLM init failed: {e}")
        _llm = None


# ══════════════════════════════════════════════════════
#  LOCAL KB HELPERS
# ══════════════════════════════════════════════════════
def derive_title(filename: str, content: str) -> str:
    match = re.search(r"^\s*#\s+(.+)$", content, re.MULTILINE)
    if match:
        return match.group(1).strip()
    name = os.path.splitext(filename)[0]
    name = re.sub(r"^\d+[-_]?", "", name)
    name = name.replace("-", " ").replace("_", " ").strip()
    return name or filename

def load_documents() -> list[dict]:
    docs = []
    if not os.path.isdir(KNOWLEDGE_PATH):
        return docs
    for root, _dirs, files in os.walk(KNOWLEDGE_PATH):
        for file in files:
            if file.lower().endswith(ALLOWED_EXTENSIONS):
                path = os.path.join(root, file)
                try:
                    with open(path, "r", encoding="utf-8") as f:
                        text = f.read()
                except Exception:
                    continue
                category = os.path.relpath(root, KNOWLEDGE_PATH)
                docs.append({
                    "file": file,
                    "title": derive_title(file, text),
                    "category": None if category == "." else category,
                    "content": text,
                })
    return docs

def chunk_text(text: str, size: int = CHUNK_SIZE_WORDS, overlap: int = CHUNK_OVERLAP_WORDS) -> list[str]:
    words = text.split()
    if not words: return []
    step = max(size - overlap, 1)
    chunks = []
    for i in range(0, len(words), step):
        chunk = " ".join(words[i:i + size])
        if chunk.strip(): chunks.append(chunk)
        if i + size >= len(words): break
    return chunks

def build_embedding_text(title: str, file: str, category: Optional[str], chunk: str) -> str:
    header = f"Title: {title}\nFile: {file}"
    if category: header += f"\nCategory: {category}"
    return f"{header}\n\n{chunk}"

def build_index():
    global _index, _chunks
    if not _use_st or _model is None or _faiss is None or _np is None: return
    documents = load_documents()
    embed_texts, metadata = [], []
    for doc in documents:
        for chunk in chunk_text(doc["content"]):
            embed_texts.append(build_embedding_text(doc["title"], doc["file"], doc["category"], chunk))
            metadata.append({
                "file": doc["file"], "title": doc["title"],
                "category": doc["category"], "content": chunk,
            })
    if not embed_texts:
        _index = None; _chunks = []; return
    with contextlib.redirect_stdout(StringIO()), contextlib.redirect_stderr(StringIO()):
        embeddings = _model.encode(embed_texts, show_progress_bar=False, normalize_embeddings=True)
    embeddings = _np.array(embeddings, dtype="float32")
    index = _faiss.IndexFlatIP(embeddings.shape[1])
    index.add(embeddings)
    _index = index
    _chunks = metadata
    print(f"[Server] indexed {len(embed_texts)} chunks from {len(documents)} files ✓")

def keyword_score(query: str, chunk: dict) -> float:
    q_tokens = set(re.findall(r"[A-Za-z0-9_]+", query.lower()))
    if not q_tokens: return 0.0
    searchable = f"{chunk['title']} {chunk['file']} {chunk['content']}"
    t_tokens = set(re.findall(r"[A-Za-z0-9_]+", searchable.lower()))
    overlap = q_tokens & t_tokens
    return len(overlap) / len(q_tokens)

def _local_search(query: str, top_k: int = 5, hybrid: bool = True, keyword_weight: float = 0.3) -> list[dict]:
    if _index is None or not _chunks: return []
    with contextlib.redirect_stdout(StringIO()), contextlib.redirect_stderr(StringIO()):
        qvec = _model.encode([query], normalize_embeddings=True)
    qvec = _np.array(qvec, dtype="float32")
    fetch_k = min(top_k * 3 if hybrid else top_k, len(_chunks))
    scores, indices = _index.search(qvec, fetch_k)
    candidates = []
    for score, idx in zip(scores[0], indices[0]):
        if idx < 0: continue
        chunk = _chunks[idx]
        final_score = float(score)
        if hybrid:
            kw = keyword_score(query, chunk)
            final_score = (1 - keyword_weight) * final_score + keyword_weight * kw
        candidates.append((final_score, chunk))
    candidates.sort(key=lambda x: x[0], reverse=True)
    return [{"score": s, **c} for s, c in candidates[:top_k]]


# ══════════════════════════════════════════════════════
#  WEB SEARCH
# ══════════════════════════════════════════════════════
def enhance_query_for_web(query: str) -> str:
    query_lower = query.lower()
    context_parts = []
    if not any(k in query_lower for k in ("jira", "jsm", "atlassian")):
        context_parts.append("Jira Service Management")
    if not any(k in query_lower for k in ("groovy", "scriptrunner", "adaptavist")):
        context_parts.append("Groovy ScriptRunner")
    if context_parts:
        return f"{' '.join(context_parts)} — {query}"
    return query

def _tavily_search(query: str, max_results: int = 3) -> list[dict]:
    payload = json.dumps({
        "api_key": TAVILY_API_KEY, "query": query, "max_results": max_results,
        "include_answer": True, "search_depth": "advanced", "include_domains": WEB_INCLUDE_DOMAINS,
    }).encode("utf-8")
    req = urllib.request.Request("https://api.tavily.com/search", data=payload, headers={"Content-Type": "application/json"}, method="POST")
    with urllib.request.urlopen(req, timeout=20) as resp:
        data = json.loads(resp.read().decode("utf-8"))
    results = []
    if data.get("answer"):
        results.append({"title": "Tavily AI Summary", "url": "", "content": data["answer"][:MAX_WEB_CONTENT_CHARS], "score": 1.0, "source": "tavily_answer"})
    for r in data.get("results", [])[:max_results]:
        results.append({"title": r.get("title", ""), "url": r.get("url", ""), "content": (r.get("content") or "")[:MAX_WEB_CONTENT_CHARS], "score": r.get("score", 0.0), "source": "tavily"})
    return results

def _jina_search(query: str, max_results: int = 3) -> list[dict]:
    url = f"https://s.jina.ai/{urllib.parse.quote(query)}"
    req = urllib.request.Request(url, headers={"Authorization": f"Bearer {JINA_API_KEY}", "Accept": "application/json", "X-Retain-Images": "none", "X-No-Cache": "true"}, method="GET")
    with urllib.request.urlopen(req, timeout=20) as resp:
        data = json.loads(resp.read().decode("utf-8"))
    results = []
    for item in data.get("data", [])[:max_results]:
        results.append({"title": item.get("title", ""), "url": item.get("url", ""), "content": (item.get("content") or "")[:MAX_WEB_CONTENT_CHARS], "score": 0.0, "source": "jina"})
    return results

def web_search(query: str, max_results: int = 3) -> list[dict]:
    if not _web_search_available: return []
    enhanced = enhance_query_for_web(query)
    primary = _web_search_provider_name or WEB_SEARCH_PROVIDER
    if primary == "tavily" and TAVILY_API_KEY:
        try: return _tavily_search(enhanced, max_results)
        except Exception as e: print(f"[Server] Tavily search failed: {e}")
    elif primary == "jina" and JINA_API_KEY:
        try: return _jina_search(enhanced, max_results)
        except Exception as e: print(f"[Server] Jina search failed: {e}")
    if primary != "jina" and JINA_API_KEY:
        try: return _jina_search(enhanced, max_results)
        except: pass
    if primary != "tavily" and TAVILY_API_KEY:
        try: return _tavily_search(enhanced, max_results)
        except: pass
    return []


# ══════════════════════════════════════════════════════
#  LLM & AGENT
# ══════════════════════════════════════════════════════
SYSTEM_PROMPT = """You are an expert assistant for Jira Service Management (JSM) administration and Groovy/ScriptRunner development on Atlassian platforms.
You receive:
1. LOCAL KNOWLEDGE BASE results — curated Groovy tutorials and ScriptRunner docs.
2. WEB SEARCH results — live results from Atlassian Community, docs, and Adaptavist.
ANSWER RULES:
• Give a direct, actionable answer with step-by-step instructions.
• Include COMPLETE, working Groovy code in ```groovy fenced blocks when scripting is involved.
• Prefer LOCAL results for Groovy/ScriptRunner syntax and patterns.
• Use WEB results for JSM configuration, newer API changes, or topics not covered locally.
• Cite sources inline as [Local: filename] or [Web: domain].
• Be concise — no filler, no repetition. Use markdown headings, code blocks, and bullet points.
• Maximum length: ~{max_tokens} tokens."""

LOCAL_CHUNK_CHARS_FOR_LLM = 800

def _format_local_for_llm(results: list[dict]) -> str:
    if not results: return "(no local results found)"
    parts = []
    for i, r in enumerate(results, 1):
        parts.append(f"[{i}] File: {r.get('file', '?')}\n    Title: {r.get('title', '?')}\n    Content:\n{r.get('content', '')[:LOCAL_CHUNK_CHARS_FOR_LLM]}")
    return "\n\n".join(parts)

def _format_web_for_llm(results: list[dict]) -> str:
    if not results: return "(no web results found)"
    parts = []
    for i, r in enumerate(results, 1):
        url = r.get("url", "")
        domain = urllib.parse.urlparse(url).netloc if url else "tavily-summary"
        parts.append(f"[{i}] Title: {r.get('title', '?')}\n    Source: {domain}\n    Content:\n{r.get('content', '')[:MAX_WEB_CONTENT_CHARS]}")
    return "\n\n".join(parts)

def synthesize_with_llm(query: str, local_results: list[dict], web_results: list[dict]) -> str:
    if _llm is None: return ""
    system = SYSTEM_PROMPT.format(max_tokens=MAX_FINAL_ANSWER_TOKENS)
    user = (f"USER QUESTION:\n{query}\n\n"
            f"=== LOCAL KNOWLEDGE BASE RESULTS ===\n{_format_local_for_llm(local_results)}\n\n"
            f"=== WEB SEARCH RESULTS ===\n{_format_web_for_llm(web_results)}\n\n"
            f"Provide a comprehensive answer based on the above sources.")
    try:
        from langchain_core.messages import SystemMessage, HumanMessage
        response = _llm.invoke([SystemMessage(content=system), HumanMessage(content=user)])
        return response.content
    except Exception as e:
        print(f"[Server] LLM synthesis failed: {e}")
        return ""

def _create_agent_executor():
    global _agent_executor
    from langchain.agents import create_tool_calling_agent, AgentExecutor
    from langchain_core.tools import tool
    from langchain_core.prompts import ChatPromptTemplate

    @tool
    def search_local_kb(query: str) -> str:
        """Search the local knowledge base of Groovy tutorials and ScriptRunner documentation. Use this FIRST for Groovy code examples."""
        results = _local_search(query, top_k=5, hybrid=True)
        return _format_local_for_llm(results) if results else "No local results found."

    @tool
    def search_web(query: str) -> str:
        """Search the web for Jira Service Management configuration or newer API docs not found locally."""
        results = web_search(query, max_results=MAX_WEB_RESULTS)
        return _format_web_for_llm(results) if results else "No web results found."

    tools = [search_local_kb, search_web]
    prompt = ChatPromptTemplate.from_messages([
        ("system", SYSTEM_PROMPT.format(max_tokens=MAX_FINAL_ANSWER_TOKENS) +
         "\n\nYou have access to tools. Use `search_local_kb` first. Use `search_web` if local is insufficient."),
        ("user", "{input}"),
        ("placeholder", "{agent_scratchpad}"),
    ])
    agent = create_tool_calling_agent(_llm, tools, prompt)
    _agent_executor = AgentExecutor(agent=agent, tools=tools, max_iterations=AGENT_MAX_ITERATIONS, verbose=False, return_intermediate_steps=False)
    print("[Server] LangChain agent created ✓")

def run_agent(query: str) -> str:
    if _agent_executor is None: return ""
    try:
        result = _agent_executor.invoke({"input": query})
        return result.get("output", "")
    except Exception as e:
        print(f"[Server] agent execution failed: {e}")
        return ""


# ══════════════════════════════════════════════════════
#  MODELS
# ══════════════════════════════════════════════════════
class SearchRequest(BaseModel):
    query: str
    top_k: int = 5
    hybrid: bool = True
    keyword_weight: float = 0.3
    
    # Unified Smart Search controls
    web_search: bool = True             # Enable web search
    max_web_results: int = 3            # Web results limit
    synthesize: bool = True             # Enable LLM synthesis
    use_agent: bool = False             # Enable multi-step LLM Agent

class SearchResultItem(BaseModel):
    file: str
    title: str
    category: Optional[str] = None
    score: float
    content: str

class WebSearchResultItem(BaseModel):
    title: str
    url: str
    content: str
    score: float = 0.0
    source: str = "web"

class SearchSource(BaseModel):
    type: str
    title: str
    file: Optional[str] = None
    url: Optional[str] = None
    score: float = 0.0

class SearchResponse(BaseModel):
    query: str
    answer: Optional[str] = None
    local_results: list[SearchResultItem] = []
    web_results: list[WebSearchResultItem] = []
    sources: list[SearchSource] = []
    used_web_search: bool = False
    used_llm: bool = False
    used_agent: bool = False
    llm_provider: Optional[str] = None
    error: Optional[str] = None

class ReloadResponse(BaseModel):
    status: str
    chunks_indexed: int

class HealthResponse(BaseModel):
    status: str
    model_loaded: bool
    faiss_loaded: bool
    chunks_indexed: int
    knowledge_path: str
    web_search_available: bool
    web_search_provider: Optional[str] = None
    llm_loaded: bool
    llm_provider: Optional[str] = None
    llm_model: Optional[str] = None
    agent_enabled: bool
    auth_enabled: bool = False   # ✅ جديد


# ══════════════════════════════════════════════════════
#  ENDPOINTS
# ══════════════════════════════════════════════════════
@app.get("/debug", dependencies=[Depends(verify_api_key)])
def debug():
    files = []
    if os.path.isdir(KNOWLEDGE_PATH):
        for root, _dirs, fs in os.walk(KNOWLEDGE_PATH):
            for f in fs: files.append(os.path.join(root, f))
    return {"path": KNOWLEDGE_PATH, "exists": os.path.exists(KNOWLEDGE_PATH), "count": len(files), "files": files[:50]}

@app.get("/health", response_model=HealthResponse)
def health():
    return {
        "status": "ok",
        "model_loaded": _use_st,
        "faiss_loaded": _faiss is not None,
        "chunks_indexed": len(_chunks),
        "knowledge_path": KNOWLEDGE_PATH,
        "web_search_available": _web_search_available,
        "web_search_provider": _web_search_provider_name,
        "llm_loaded": _llm is not None,
        "llm_provider": _llm_provider_name,
        "llm_model": LLM_MODEL if _llm is not None else None,
        "agent_enabled": _agent_executor is not None,
        "auth_enabled": AUTH_ENABLED,  # ✅ جديد
    }

@app.post("/reload", response_model=ReloadResponse, dependencies=[Depends(verify_api_key)])
def reload_knowledge_base():
    build_index()
    return ReloadResponse(status="ok", chunks_indexed=len(_chunks))

@app.get("/ping")
def ping():
    return {"pong": True}


# ══════════════════════════════════════════════════════
#  UNIFIED SEARCH ENDPOINT
# ══════════════════════════════════════════════════════
@app.post("/search", response_model=SearchResponse, dependencies=[Depends(verify_api_key)])
def search(req: SearchRequest):
    """
    Unified Intelligent Search endpoint.
    
    Performs:
    1. Local FAISS semantic search.
    2. Web search (Tavily/Jina) if `web_search=True`.
    3. LLM Synthesis (Cerebras/Cohere) if `synthesize=True`.
    4. Multi-step Agent reasoning if `use_agent=True`.
    """
    response = SearchResponse(query=req.query)
    errors = []

    # ── Agent Mode ──
    if req.use_agent and _agent_executor is not None:
        answer = run_agent(req.query)
        if answer:
            response.answer = answer
            response.used_llm = True
            response.used_agent = True
            response.llm_provider = _llm_provider_name
            return response
        else:
            errors.append("Agent execution failed — falling back to RAG mode")

    # ── RAG Mode ──

    # 1) Local KB Search
    local_raw: list[dict] = []
    if _use_st and _model is not None and _index is not None and _chunks:
        local_raw = _local_search(req.query, req.top_k, req.hybrid, req.keyword_weight)
        response.local_results = [
            SearchResultItem(
                file=r["file"], title=r["title"], category=r["category"],
                score=r["score"], content=r["content"],
            ) for r in local_raw
        ]

    # 2) Web Search
    web_raw: list[dict] = []
    if req.web_search and _web_search_available:
        try:
            web_raw = web_search(req.query, req.max_web_results)
            response.web_results = [
                WebSearchResultItem(
                    title=r.get("title", ""),
                    url=r.get("url", ""),
                    content=r.get("content", ""),
                    score=r.get("score", 0.0),
                    source=r.get("source", "web"),
                ) for r in web_raw
            ]
            response.used_web_search = True
        except Exception as e:
            errors.append(f"Web search error: {e}")

    # 3) Consolidate Sources
    for r in local_raw:
        response.sources.append(SearchSource(
            type="local",
            title=r.get("title", ""),
            file=r.get("file", ""),
            score=r.get("score", 0.0),
        ))
    for r in web_raw:
        response.sources.append(SearchSource(
            type="web",
            title=r.get("title", ""),
            url=r.get("url", ""),
            score=r.get("score", 0.0),
        ))

    # 4) LLM Synthesis
    if req.synthesize and _llm is not None and (local_raw or web_raw):
        answer = synthesize_with_llm(req.query, local_raw, web_raw)
        if answer:
            response.answer = answer
            response.used_llm = True
            response.llm_provider = _llm_provider_name
        else:
            errors.append("LLM synthesis returned empty")
    elif req.synthesize and _llm is None:
        errors.append("LLM not configured — returning raw results without synthesis")

    if errors:
        response.error = "; ".join(errors)

    return response