@skyphusion/search-mcp 0.5.1 → 0.7.0

package/README.md

@@ -89,6 +89,23 @@ wrangler secret put MCP_TOKEN -c wrangler.mcp.toml
 wrangler secret put TURNSTILE_SECRET   # optional; skips verification when unset
 ```
 
+### Service binding RPC
+
+`POST /ask` stays the browser route (CORS, Turnstile, per-IP rate limit). A Worker on the same account can retrieve chunks without that route by binding to the named entrypoint `SearchEntrypoint`, exported from `src/index.ts` (the public query Worker and the rockenhaus Worker, which share that file). The MCP Worker does not export it.
+
+The entrypoint is not a URL. Internet `fetch` still hits the default export only. The caller selects the class in its own wrangler config. This Worker's wrangler bindings do not change. `entrypoint` is the caller's service-binding field ([RPC named entrypoints](https://developers.cloudflare.com/workers/runtime-apis/bindings/service-bindings/rpc/#named-entrypoints)):
+
+```toml
+[[services]]
+binding = "SEARCH_RPC"
+service = "skyphusion-search"   # or rockenhaus-search; a self-host uses its own script name
+entrypoint = "SearchEntrypoint"
+```
+
+`await env.SEARCH_RPC.search(question)` returns `{ chunks }` from the same `SEARCH` binding `/ask` uses. The retrieval options match `/ask` (hybrid, 6 results, query rewrite on). Nothing is generated: the call is `SEARCH.search`, not `chatCompletions`. The question must be a non-empty string of at most 2000 characters after trim, the same bound as `/ask`. A bad call throws `Error` with one of these messages: `empty_question`, `question_too_long`, `rate_limited`, `search_unavailable`.
+
+`ASK_LIMITER` still applies. The key is the fixed string `ask:rpc` for every binding caller. There is no Origin check and no Turnstile check: the method has no HTTP request, and only a Worker whose config names this entrypoint receives a stub.
+
 ### MCP client wiring
 
 ```json

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@skyphusion/search-mcp",
-  "version": "0.5.1",
+  "version": "0.7.0",
   "description": "Bearer-gated MCP server and streaming query Worker over Cloudflare AI Search, plus a git-to-R2 corpus sync toolkit.",
   "license": "AGPL-3.0-only",
   "author": "Conrad Rockenhaus <conrad@skyphusion.org>",
@@ -72,12 +72,13 @@
   "devDependencies": {
     "@cloudflare/workers-types": "^5.20260713.1",
     "@types/node": "^26.2.0",
-    "@vitest/coverage-istanbul": "^4.1.10",
+    "@vitest/coverage-istanbul": "^5.0.3",
     "typescript": "^7.0.2",
-    "vitest": "^4.1.10",
+    "vitest": "^5.0.3",
     "wrangler": "^4.110.0"
   },
   "overrides": {
-    "undici": "7.29.0"
+    "undici": "7.29.0",
+    "sharp": "0.35.4"
   }
 }

package/public/ask-widget.js

@@ -193,18 +193,40 @@
     }
 
     function handleEvent(block) {
-      var isChunks = /(^|\n)event:\s*chunks/.test(block);
-      var m = block.match(/(^|\n)data:\s?(.*)$/s);
-      if (!m) return;
-      var payload = m[2].trim();
+      // One SSE event block: an optional "event:" name plus one or more "data:" lines,
+      // which the SSE spec joins with "\n". An error event throws so the request's
+      // catch shows it: an upstream failure must never read as an empty corpus.
+      var name = "";
+      var lines = [];
+      block.split("\n").forEach(function (line) {
+        var f = line.match(/^(event|data):\s?(.*)$/);
+        if (!f) return;
+        if (f[1] === "event") name = f[2].trim();
+        else lines.push(f[2]);
+      });
+      if (!lines.length) {
+        if (name === "error") throw new Error("stream error");
+        return;
+      }
+      var payload = lines.join("\n").trim();
       if (payload === "[DONE]") return;
       var data;
       try {
         data = JSON.parse(payload);
       } catch (e) {
+        if (name === "error") throw new Error("stream error");
         return;
       }
-      if (isChunks) {
+      if (name === "error" || (data && data.error)) {
+        var err = data && data.error;
+        throw new Error(
+          (typeof err === "string" && err) ||
+            (err && err.message) ||
+            (data && data.message) ||
+            "stream error",
+        );
+      }
+      if (name === "chunks") {
         renderSources(data);
         return;
       }
@@ -246,7 +268,12 @@
           var buf = "";
           function pump() {
             return reader.read().then(function (r) {
-              if (r.done) return;
+              if (r.done) {
+                // Flush what is left: the final event may lack its trailing blank line.
+                buf += decoder.decode();
+                if (buf.trim()) handleEvent(buf);
+                return;
+              }
               buf += decoder.decode(r.value, { stream: true });
               var parts = buf.split("\n\n");
               buf = parts.pop();