clef-mcp 0.1.4 → 0.1.5

package/dist/cli.js

@@ -853,6 +853,7 @@ var LlamaCppRuntime = class {
     this.host = opts.host ?? "127.0.0.1";
     this.loadTimeoutMs = opts.loadTimeoutMs ?? 12e4;
     this.requestTimeoutMs = opts.requestTimeoutMs ?? 6e4;
+    this.batchSize = opts.batchSize ?? (Number.parseInt(process.env.CLEF_LLAMA_BATCH ?? "", 10) || 8192);
   }
   opts;
   proc;
@@ -862,6 +863,7 @@ var LlamaCppRuntime = class {
   requestTimeoutMs;
   /** Serializes decide() calls: the llama.cpp clef path handles one sequence per batch. */
   queue = Promise.resolve();
+  batchSize;
   get endpoint() {
     if (!this.resolvedPort) throw new ClefError(ClefErrorCode.RUNTIME_INIT_FAILED, "Runtime is not started");
     return `http://${this.host}:${this.resolvedPort}`;
@@ -881,6 +883,10 @@ var LlamaCppRuntime = class {
       "--port",
       String(port),
       "--no-webui",
+      "-b",
+      String(this.batchSize),
+      "-ub",
+      String(this.batchSize),
       ...this.opts.alias ? ["--alias", this.opts.alias] : []
     ];
     this.opts.log?.debug("spawning llama-server", { bin: this.opts.serverBin, args });
@@ -998,7 +1004,8 @@ ${tail}` : ""),
       }
       throw new ClefError(
         ClefErrorCode.CLEF_INFERENCE_FAILED,
-        `Inference failed with HTTP ${res.status}: ${apiMessage.slice(0, 2e3)}`
+        `Inference failed with HTTP ${res.status}: ${apiMessage.slice(0, 2e3)}`,
+        /physical batch size/i.test(apiMessage) ? "Too many/too large questions for one request. Split into several calls, reduce criteria text, or raise the batch with CLEF_LLAMA_BATCH." : void 0
       );
     }
     if (!body || typeof body !== "object") {
@@ -1982,7 +1989,7 @@ async function removeManagedRuntime(cfg) {
 init_prompt();
 
 // src/version.ts
-var VERSION = "0.1.4";
+var VERSION = "0.1.5";
 
 // src/cli/install.ts
 var OS_NAMES = {