3 -> 2 candidates

mcp/server.py

@@ -59,14 +59,22 @@ def embed(text, input_type):
     assert resp.embeddings.float_ is not None
     return resp.embeddings.float_[0]
 
-def search(query, roles: list[str], top_n: int = 3, max_content_length: int = 1500) -> list[dict]:
-    """Search with vector similarity, then rerank with Cohere for better relevance."""
+@mcp.tool
+def search_caving_documents(query: str, priority_prefixes: list[str] | None = None) -> dict:
+    """Search caving documents for information about caves, techniques, safety, accidents, history, and more.
+
+    Args:
+        query: Search query
+        priority_prefixes: Optional list of key prefixes to prioritize (e.g., ['nss/aca'] for rescue topics)
+    """
+    roles = get_user_roles()
+    if not roles:
+        return {"results": [], "note": "No results. Answer based on your knowledge."}
+
     query_embedding = embed(query, 'search_query')
 
-    if not roles:
-        return []
-
     # Fetch more candidates for reranking
+    top_n = 2
     candidate_limit = top_n * 4
     rows = conn.execute(
         'SELECT * FROM embeddings WHERE embedding IS NOT NULL AND role = ANY(%s) ORDER BY embedding <=> %s::vector LIMIT %s',
@@ -74,52 +82,37 @@ def search(query, roles: list[str], top_n: int = 3, max_content_length: int = 15
     ).fetchall()
 
     if not rows:
-        return []
+        return {"results": [], "note": "No results found. Answer based on your knowledge."}
 
     # Rerank with Cohere for better relevance
     rerank_resp = co.rerank(
         query=query,
         documents=[row['content'] or '' for row in rows],
         model='rerank-v3.5',
-        top_n=top_n,
+        top_n=min(top_n * 2, len(rows)),
     )
 
+    # Build results with optional priority boost
     docs = []
     for result in rerank_resp.results:
         row = rows[result.index]
+        score = result.relevance_score
+
+        # Boost score if key starts with any priority prefix (e.g., 'nss/aca')
+        if priority_prefixes:
+            key = row['key'] or ''
+            if any(key.startswith(prefix) for prefix in priority_prefixes):
+                score = min(1.0, score * 1.3)
+
         content = row['content'] or ''
-        if len(content) > max_content_length:
-            content = content[:max_content_length] + '...[truncated, use get_document_page for full text]'
-        docs.append({'key': row['key'], 'content': content, 'relevance': round(result.relevance_score, 3)})
-    return docs
+        docs.append({'key': row['key'], 'content': content, 'relevance': round(score, 3)})
 
-@mcp.tool
-def get_cave_location(cave: str, state: str, county: str) -> list[dict]:
-    """Lookup cave location as coordinates."""
-    roles = get_user_roles()
-    return search(f'{cave} Location, latitude, Longitude. Located in {state} and {county} county.', roles)
-
-@mcp.tool
-def general_caving_information(query: str) -> list[dict]:
-    """General purpose search for any topic related to caves."""
-    roles = get_user_roles()
-    return search(query, roles)
-
-@mcp.tool
-def get_document_page(key: str) -> dict:
-    """Fetch full content for a document page. Pass the exact 'key' value from search results."""
-    roles = get_user_roles()
-    if not roles:
-        return {"error": "No roles assigned"}
-
-    row = conn.execute(
-        'SELECT key, content FROM embeddings WHERE key = %s AND role = ANY(%s)',
-        (key, roles)
-    ).fetchone()
-
-    if row:
-        return {"key": row["key"], "content": row["content"]}
-    return {"error": f"Page not found: {key}"}
+    # Re-sort by boosted score and return top_n
+    docs.sort(key=lambda x: x['relevance'], reverse=True)
+    return {
+        "results": docs[:top_n],
+        "note": "These are ALL available results. Do NOT search again - answer using these results now."
+    }
 
 @mcp.tool
 def get_user_info() -> dict:

web/agent/src/agent.py

@@ -26,7 +26,6 @@ logfire.configure(
 logfire.instrument_pydantic_ai()
 logfire.instrument_httpx()
 
-from typing import Any
 from pydantic_ai import Agent, ModelMessage, RunContext
 from pydantic_ai.settings import ModelSettings
 from pydantic_ai.mcp import CallToolFunc
@@ -43,8 +42,8 @@ def limit_history(ctx: RunContext[None], messages: list[ModelMessage]) -> list[M
     if not messages:
         return messages
 
-    # Keep only the last 4 messages
-    messages = messages[-4:]
+    # Keep last 10 messages
+    messages = messages[-10:]
 
     # Check if the last message is an assistant response with a tool call
     # If so, remove it - it's orphaned (no tool result followed)
@@ -80,37 +79,34 @@ AGENT_INSTRUCTIONS = """Caving assistant. Help with exploration, safety, surveyi
 Rules:
 1. ALWAYS cite sources in a bulleted list at the end of every reply, even if there's only one. Format them human-readably (e.g., "- The Trog 2021, page 19" not "vpi/trog/2021-trog.pdf/page-19.pdf").
 2. Say when uncertain. Never hallucinate.
-3. Be safety-conscious.
-4. Can create ascii diagrams/maps.
-5. Be direct—no sycophantic phrases.
-6. Keep responses concise.
-7. Use tools sparingly—one search usually suffices.
-8. If you hit the search limit, end your reply with an italicized note: *Your question may be too broad. Try asking something more specific.* Do NOT mention "tools" or "tool limits"—the user doesn't know what those are."""
+3. Be direct—no sycophantic phrases.
+4. Keep responses concise.
+5. SEARCH EXACTLY ONCE. After searching, IMMEDIATELY answer using those results. NEVER search again - additional searches are blocked and waste resources.
+6. For rescue, accident, or emergency-related queries, use priority_prefixes=['nss/aca'] when searching to prioritize official accident reports."""
+
+SOURCES_ONLY_INSTRUCTIONS = """SOURCES ONLY MODE: Give exactly ONE sentence summary. Then list sources with specific page numbers (e.g., "- The Trog 2021, page 19"). No explanations."""
 
 
-def create_tool_call_limiter(max_calls: int = 3):
-    """Create a process_tool_call callback that limits tool calls."""
-    call_count = [0]  # Mutable container for closure
+def create_search_limiter():
+    """Block searches after the first one."""
+    searched = [False]
 
     async def process_tool_call(
         ctx: RunContext,
         call_tool: CallToolFunc,
         name: str,
-        tool_args: dict[str, Any],
+        tool_args: dict,
     ):
-        call_count[0] += 1
-        if call_count[0] > max_calls:
-            return (
-                f"SEARCH LIMIT REACHED: You have made {max_calls} searches. "
-                "Stop searching and answer now with what you have. "
-                "End your reply with: *Your question may be too broad. Try asking something more specific.*"
-            )
+        if name == "search_caving_documents":
+            if searched[0]:
+                return "You have already searched. Use the results you have."
+            searched[0] = True
         return await call_tool(name, tool_args)
 
     return process_tool_call
 
 
-def create_agent(user_roles: list[str] | None = None):
+def create_agent(user_roles: list[str] | None = None, sources_only: bool = False):
     """Create an agent with MCP tools configured for the given user roles."""
     toolsets = []
 
@@ -129,7 +125,7 @@ def create_agent(user_roles: list[str] | None = None):
                 url=CAVE_MCP_URL,
                 headers={"x-user-roles": roles_header},
                 timeout=30.0,
-                process_tool_call=create_tool_call_limiter(max_calls=3),
+                process_tool_call=create_search_limiter(),
             )
             toolsets.append(mcp_server)
             logger.info(f"MCP server configured with roles: {user_roles}")
@@ -140,10 +136,15 @@ def create_agent(user_roles: list[str] | None = None):
     else:
         logger.info("MCP server unavailable - running without MCP tools")
 
+    # Build instructions based on mode
+    instructions = AGENT_INSTRUCTIONS
+    if sources_only:
+        instructions = f"{SOURCES_ONLY_INSTRUCTIONS}\n\n{AGENT_INSTRUCTIONS}"
+
     return Agent(
         model="anthropic:claude-sonnet-4-5",
         toolsets=toolsets if toolsets else None,
-        instructions=AGENT_INSTRUCTIONS,
+        instructions=instructions,
         history_processors=[limit_history],
         model_settings=ModelSettings(max_tokens=4096),
     )

web/agent/src/main.py

@@ -67,8 +67,13 @@ async def handle_agent_request(request: Request) -> Response:
         except json.JSONDecodeError as e:
             logger.warning(f"Failed to parse x-user-roles header: {e}")
 
-    # Create agent with the user's roles
-    agent = create_agent(user_roles)
+    # Extract sources-only mode from header
+    sources_only = request.headers.get("x-sources-only", "false") == "true"
+    if sources_only:
+        logger.info("Sources-only mode enabled")
+
+    # Create agent with the user's roles and mode
+    agent = create_agent(user_roles, sources_only=sources_only)
 
     # Dispatch the request - tool limits handled by ToolCallLimiter in agent.py
     return await AGUIAdapter.dispatch_request(

web/src/app/api/copilotkit/route.ts

@@ -15,13 +15,19 @@ export const POST = async (req: NextRequest) => {
   const session = await auth0.getSession();
   const userRoles = (session?.user?.roles as string[]) || [];
 
-  console.log("DEBUG: User roles from session:", userRoles);
+  // Get sources-only mode from query param
+  const url = new URL(req.url);
+  const sourcesOnly = url.searchParams.get("sourcesOnly") === "true";
 
-  // Create HttpAgent with user roles header
+  console.log("DEBUG: User roles from session:", userRoles);
+  console.log("DEBUG: Sources only mode:", sourcesOnly);
+
+  // Create HttpAgent with user roles and sources-only headers
   const agent = new HttpAgent({
     url: process.env.AGENT_URL || "http://localhost:8000/",
     headers: {
       "x-user-roles": JSON.stringify(userRoles),
+      "x-sources-only": sourcesOnly ? "true" : "false",
     },
   });
 

web/src/app/layout.tsx

@@ -1,6 +1,5 @@
 import type { Metadata } from "next";
 
-import { CopilotKit } from "@copilotkit/react-core";
 import { Auth0Provider } from "@auth0/nextjs-auth0/client";
 import "./globals.css";
 import "@copilotkit/react-ui/styles.css";
@@ -19,9 +18,7 @@ export default function RootLayout({
     <html lang="en">
       <body className={"antialiased"}>
         <Auth0Provider>
-          <CopilotKit runtimeUrl="/api/copilotkit" agent="vpi_1000">
           {children}
-          </CopilotKit>
         </Auth0Provider>
       </body>
     </html>

web/src/app/page.tsx

@@ -1,6 +1,6 @@
 "use client";
 
-import { useCopilotAction, useCopilotChat } from "@copilotkit/react-core";
+import { CopilotKit, useCopilotAction, useCopilotChat } from "@copilotkit/react-core";
 import { CopilotKitCSSProperties, CopilotChat } from "@copilotkit/react-ui";
 import { useState } from "react";
 import { useUser } from "@auth0/nextjs-auth0/client";
@@ -36,10 +36,8 @@ function LoadingOverlay() {
   }
 }
 
-export default function CopilotKitPage() {
-  const [themeColor, setThemeColor] = useState("#6366f1");
-  const { user, isLoading: authLoading } = useUser();
-
+// Chat content with CopilotKit hooks - must be inside CopilotKit provider
+function ChatContent({ themeColor, setThemeColor }: { themeColor: string; setThemeColor: (color: string) => void }) {
   useCopilotAction({
     name: "setThemeColor",
     parameters: [{
@@ -52,6 +50,35 @@ export default function CopilotKitPage() {
     },
   });
 
+  return (
+    <div
+      className="flex-1 flex justify-center py-8 px-2 overflow-hidden relative"
+      style={{ "--copilot-kit-primary-color": themeColor } as CopilotKitCSSProperties}
+    >
+      <div className="h-full w-full max-w-5xl flex flex-col">
+        <CopilotChat
+          labels={{
+            title: "AI Cartwright",
+            initial: "Hello! I'm here to help with anything related to caving. Ask me about caves, techniques, safety, equipment, or anything else caving-related!",
+          }}
+          className="h-full w-full"
+        />
+      </div>
+      <LoadingOverlay />
+    </div>
+  );
+}
+
+export default function CopilotKitPage() {
+  const [themeColor, setThemeColor] = useState("#6366f1");
+  const [sourcesOnlyMode, setSourcesOnlyMode] = useState(false);
+  const { user, isLoading: authLoading } = useUser();
+
+  // Dynamic runtime URL based on sources-only mode
+  const runtimeUrl = sourcesOnlyMode
+    ? "/api/copilotkit?sourcesOnly=true"
+    : "/api/copilotkit";
+
   // Show loading state while checking authentication
   if (authLoading) {
     return (
@@ -88,15 +115,26 @@ export default function CopilotKitPage() {
 
   // If authenticated, show the CopilotKit chat with user profile
   return (
-    <main
-      style={{ "--copilot-kit-primary-color": themeColor } as CopilotKitCSSProperties}
-      className="h-screen w-screen flex flex-col bg-gray-50"
+    <CopilotKit
+      runtimeUrl={runtimeUrl}
+      agent="vpi_1000"
+      key={sourcesOnlyMode ? "sources" : "normal"}
     >
+      <main className="h-screen w-screen flex flex-col bg-gray-50">
         {/* Header with user profile and logout */}
         <div className="w-full bg-white shadow-sm border-b border-gray-200 px-4 py-3">
           <div className="max-w-7xl mx-auto flex justify-between items-center">
             <div className="flex items-center gap-4">
               <h1 className="text-xl font-semibold text-gray-900">Cavepedia</h1>
+              <label className="flex items-center gap-2 text-sm text-gray-600 cursor-pointer">
+                <input
+                  type="checkbox"
+                  checked={sourcesOnlyMode}
+                  onChange={(e) => setSourcesOnlyMode(e.target.checked)}
+                  className="w-4 h-4 rounded border-gray-300 text-indigo-600 focus:ring-indigo-500"
+                />
+                Sources only
+              </label>
             </div>
             <div className="flex items-center gap-4">
               {user.picture && (
@@ -120,18 +158,8 @@ export default function CopilotKitPage() {
         </div>
 
         {/* CopilotKit Chat */}
-      <div className="flex-1 flex justify-center py-8 px-2 overflow-hidden relative">
-        <div className="h-full w-full max-w-5xl flex flex-col">
-          <CopilotChat
-            labels={{
-              title: "AI Cartwright",
-              initial: "Hello! I'm here to help with anything related to caving. Ask me about caves, techniques, safety, equipment, or anything else caving-related!",
-            }}
-            className="h-full w-full"
-          />
-        </div>
-        <LoadingOverlay />
-      </div>
+        <ChatContent themeColor={themeColor} setThemeColor={setThemeColor} />
       </main>
+    </CopilotKit>
   );
 }