@atlaskit/editor-plugin-autocomplete 2.4.0 → 2.5.0

package/dist/esm/pm-plugins/local-slow-lane-client.js

@@ -0,0 +1,439 @@
+import _slicedToArray from "@babel/runtime/helpers/slicedToArray";
+import _toConsumableArray from "@babel/runtime/helpers/toConsumableArray";
+import _defineProperty from "@babel/runtime/helpers/defineProperty";
+import _asyncToGenerator from "@babel/runtime/helpers/asyncToGenerator";
+function _createForOfIteratorHelper(r, e) { var t = "undefined" != typeof Symbol && r[Symbol.iterator] || r["@@iterator"]; if (!t) { if (Array.isArray(r) || (t = _unsupportedIterableToArray(r)) || e && r && "number" == typeof r.length) { t && (r = t); var _n = 0, F = function F() {}; return { s: F, n: function n() { return _n >= r.length ? { done: !0 } : { done: !1, value: r[_n++] }; }, e: function e(r) { throw r; }, f: F }; } throw new TypeError("Invalid attempt to iterate non-iterable instance.\nIn order to be iterable, non-array objects must have a [Symbol.iterator]() method."); } var o, a = !0, u = !1; return { s: function s() { t = t.call(r); }, n: function n() { var r = t.next(); return a = r.done, r; }, e: function e(r) { u = !0, o = r; }, f: function f() { try { a || null == t.return || t.return(); } finally { if (u) throw o; } } }; }
+function _unsupportedIterableToArray(r, a) { if (r) { if ("string" == typeof r) return _arrayLikeToArray(r, a); var t = {}.toString.call(r).slice(8, -1); return "Object" === t && r.constructor && (t = r.constructor.name), "Map" === t || "Set" === t ? Array.from(r) : "Arguments" === t || /^(?:Ui|I)nt(?:8|16|32)(?:Clamped)?Array$/.test(t) ? _arrayLikeToArray(r, a) : void 0; } }
+function _arrayLikeToArray(r, a) { (null == a || a > r.length) && (a = r.length); for (var e = 0, n = Array(a); e < a; e++) n[e] = r[e]; return n; }
+import _regeneratorRuntime from "@babel/runtime/regenerator";
+function ownKeys(e, r) { var t = Object.keys(e); if (Object.getOwnPropertySymbols) { var o = Object.getOwnPropertySymbols(e); r && (o = o.filter(function (r) { return Object.getOwnPropertyDescriptor(e, r).enumerable; })), t.push.apply(t, o); } return t; }
+function _objectSpread(e) { for (var r = 1; r < arguments.length; r++) { var t = null != arguments[r] ? arguments[r] : {}; r % 2 ? ownKeys(Object(t), !0).forEach(function (r) { _defineProperty(e, r, t[r]); }) : Object.getOwnPropertyDescriptors ? Object.defineProperties(e, Object.getOwnPropertyDescriptors(t)) : ownKeys(Object(t)).forEach(function (r) { Object.defineProperty(e, r, Object.getOwnPropertyDescriptor(t, r)); }); } return e; }
+/**
+ * Local Slow Lane Client: On-device inference via @mlc-ai/web-llm.
+ *
+ * Drop-in replacement for the network-based slow-lane-client.  Instead of
+ * calling a backend API, this client uses MLC WebLLM to run a small language
+ * model (SmolLM 135M) directly in the browser via WebGPU.
+ *
+ * ── Why main thread (no Web Worker)? ─────────────────────────────────────
+ * SmolLM 135M is small enough (~270 MB weights, 350-400 MB VRAM) that
+ * WebGPU inference on the main thread is production-viable:
+ *
+ *   - WebGPU GPU compute is inherently async (doesn't block the main thread)
+ *   - CPU overhead (tokenization + post-processing) is only 5-10 ms
+ *   - Single forward pass latency is 50-150 ms — well within autocomplete
+ *     expectations (~250 ms between word boundaries)
+ *
+ * This avoids all the complexity of Web Workers:
+ *   - No CSP workarounds (blob URLs, inline scripts)
+ *   - No bundler configuration (worker-plugin, import.meta.url)
+ *   - No message passing protocol
+ *   - Standard npm import — just works
+ *
+ * ── Interface ────────────────────────────────────────────────────────────
+ * Same shape as createSlowLaneClient so text-predictor.ts needs zero changes.
+ * The client exposes getContextVector() and getLmLogits() which are populated
+ * asynchronously after each updateContext() call.
+ */
+
+import { isAutocompleteDebugEnabled } from './debug-mode';
+import { isWordBoundary } from './slow-lane-client';
+
+// ─── Types ───────────────────────────────────────────────────────────────────
+
+// Same return type as createSlowLaneClient for drop-in compatibility
+
+// ─── Constants ───────────────────────────────────────────────────────────────
+
+var DEFAULT_DEBOUNCE_MS = 300;
+export var LOCAL_MLC_MODEL_ID = 'SmolLM2-135M-Instruct-q0f16-MLC';
+
+/** HF root for the default weights (includes `tensor-cache.json` for WebLLM 0.2+). */
+export var LOCAL_MLC_HF_MODEL_REPO = 'https://huggingface.co/mlc-ai/SmolLM2-135M-Instruct-q0f16-MLC';
+export var LOCAL_MLC_MODEL_LIB_WASM_NAME = 'SmolLM2-135M-Instruct-q0f16-ctx4k_cs1k-webgpu.wasm';
+
+/**
+ * Original target repo (add-basics fine-tune). **Not compatible with WebLLM 0.2.x** (no `tensor-cache.json`).
+ * @see module doc above
+ */
+export var HUGGINGFACE_TB_SMOLLM_ADD_BASICS_REPO = 'https://huggingface.co/HuggingFaceTB/smollm-135M-instruct-add-basics-q0f16-MLC';
+
+// ─── Factory ─────────────────────────────────────────────────────────────────
+
+/**
+ * Create a local slow-lane client powered by MLC WebLLM.
+ *
+ * The engine is initialised lazily — model weights are downloaded (and cached
+ * in IndexedDB) on first use. Subsequent page loads skip the download.
+ *
+ * Usage:
+ * ```ts
+ * const client = createLocalSlowLaneClient({ debounceMs: 300 });
+ * // On word boundaries:
+ * client.updateContext(docText);
+ * // In scoring pipeline:
+ * const vec = client.getContextVector();
+ * const logits = client.getLmLogits();
+ * // On plugin teardown:
+ * client.destroy();
+ * ```
+ */
+export var createLocalSlowLaneClient = function createLocalSlowLaneClient() {
+  var config = arguments.length > 0 && arguments[0] !== undefined ? arguments[0] : {};
+  var _config$debounceMs = config.debounceMs,
+    debounceMs = _config$debounceMs === void 0 ? DEFAULT_DEBOUNCE_MS : _config$debounceMs,
+    onUpdate = config.onUpdate,
+    onStatus = config.onStatus,
+    _config$modelId = config.modelId,
+    modelId = _config$modelId === void 0 ? LOCAL_MLC_MODEL_ID : _config$modelId,
+    customModelConfig = config.customModelConfig;
+
+  // ── State ──────────────────────────────────────────────────────────────
+  var storedContextVector = null;
+  var storedLmLogits = null;
+  var debounceTimer = null;
+  var lastRequestedText = '';
+  var requestCounter = 0;
+  var latestRequestId = -1;
+  var ready = false;
+  var destroyed = false;
+  var initFailed = false;
+  var engine = null;
+  var engineInitPromise = null;
+  var unloadEngine = function unloadEngine(engineToUnload) {
+    engineToUnload.unload().catch(function (error) {
+      if (isAutocompleteDebugEnabled()) {
+        // eslint-disable-next-line no-console
+        console.log('%c[LocalSlowLane] %cFailed to unload engine', 'color: #9c27b0; font-weight: bold;', 'color: inherit;', error);
+      }
+    });
+  };
+
+  // ── Engine initialisation ──────────────────────────────────────────────
+
+  var initProgressCallback = function initProgressCallback(progress) {
+    var message = "[".concat((progress.progress * 100).toFixed(0), "%] ").concat(progress.text);
+    if (isAutocompleteDebugEnabled()) {
+      // eslint-disable-next-line no-console
+      console.log("%c[LocalSlowLane] %c\uD83D\uDD04 ".concat(message), 'color: #9c27b0; font-weight: bold;', 'color: inherit;');
+    }
+    onStatus === null || onStatus === void 0 || onStatus(message);
+  };
+  var initEngine = /*#__PURE__*/function () {
+    var _ref = _asyncToGenerator( /*#__PURE__*/_regeneratorRuntime.mark(function _callee() {
+      var _yield$import, CreateMLCEngine, prebuiltAppConfig, customModelRecord, appConfig, errorMsg;
+      return _regeneratorRuntime.wrap(function _callee$(_context) {
+        while (1) switch (_context.prev = _context.next) {
+          case 0:
+            _context.prev = 0;
+            if (isAutocompleteDebugEnabled()) {
+              // eslint-disable-next-line no-console
+              console.log("%c[LocalSlowLane] %c\uD83D\uDE80 Initialising MLC engine with model: ".concat(modelId), 'color: #9c27b0; font-weight: bold;', 'color: inherit;');
+            }
+            onStatus === null || onStatus === void 0 || onStatus("Initialising model: ".concat(modelId, "\u2026"));
+            if ('gpu' in navigator) {
+              _context.next = 5;
+              break;
+            }
+            throw new Error('WebGPU not supported');
+          case 5:
+            _context.next = 7;
+            return import( /* webpackChunkName: "@atlaskit-internal_editor-plugin-autocomplete-mlc-web-llm" */'@mlc-ai/web-llm');
+          case 7:
+            _yield$import = _context.sent;
+            CreateMLCEngine = _yield$import.CreateMLCEngine;
+            prebuiltAppConfig = _yield$import.prebuiltAppConfig;
+            customModelRecord = customModelConfig ? _objectSpread(_objectSpread({
+              model: customModelConfig.model,
+              model_id: modelId,
+              model_lib: customModelConfig.modelLib,
+              low_resource_required: true,
+              required_features: ['shader-f16']
+            }, customModelConfig.vramRequiredMB !== undefined ? {
+              vram_required_MB: customModelConfig.vramRequiredMB
+            } : {}), customModelConfig.contextWindowSize !== undefined ? {
+              overrides: {
+                context_window_size: customModelConfig.contextWindowSize
+              }
+            } : {}) : undefined;
+            appConfig = {
+              model_list: [].concat(_toConsumableArray(prebuiltAppConfig.model_list), _toConsumableArray(customModelRecord ? [customModelRecord] : []))
+            };
+            _context.next = 14;
+            return CreateMLCEngine(modelId, {
+              appConfig: appConfig,
+              initProgressCallback: initProgressCallback
+            });
+          case 14:
+            engine = _context.sent;
+            if (!destroyed) {
+              _context.next = 19;
+              break;
+            }
+            // destroy() was called while we were loading — clean up
+            unloadEngine(engine);
+            engine = null;
+            return _context.abrupt("return");
+          case 19:
+            ready = true;
+            if (isAutocompleteDebugEnabled()) {
+              // eslint-disable-next-line no-console
+              console.log('%c[LocalSlowLane] %c✅ MLC engine loaded and ready', 'color: #9c27b0; font-weight: bold;', 'color: #4caf50;');
+            }
+            onStatus === null || onStatus === void 0 || onStatus('Model loaded and ready.');
+            _context.next = 32;
+            break;
+          case 24:
+            _context.prev = 24;
+            _context.t0 = _context["catch"](0);
+            errorMsg = _context.t0 instanceof Error ? _context.t0.message : String(_context.t0);
+            ready = false;
+            if (isAutocompleteDebugEnabled()) {
+              // eslint-disable-next-line no-console
+              console.log("[LocalSlowLane] Engine initialisation failed: ".concat(errorMsg));
+            }
+            onStatus === null || onStatus === void 0 || onStatus("Engine initialisation failed: ".concat(errorMsg));
+            engineInitPromise = null;
+            initFailed = true;
+          case 32:
+          case "end":
+            return _context.stop();
+        }
+      }, _callee, null, [[0, 24]]);
+    }));
+    return function initEngine() {
+      return _ref.apply(this, arguments);
+    };
+  }();
+  var ensureEngineInitialized = function ensureEngineInitialized() {
+    if (initFailed) {
+      return Promise.resolve();
+    }
+    if (!engineInitPromise) {
+      engineInitPromise = initEngine();
+    }
+    return engineInitPromise;
+  };
+
+  // ── Inference ──────────────────────────────────────────────────────────
+
+  /**
+   * Run a single forward pass to extract next-token logit probabilities.
+   *
+   * We use the chat completions API with `max_tokens: 1` and `logprobs: true`
+   * to get the model's next-token distribution without generating text.
+   * This is the cheapest possible inference call — a single forward pass.
+   */
+  var runInference = /*#__PURE__*/function () {
+    var _ref2 = _asyncToGenerator( /*#__PURE__*/_regeneratorRuntime.mark(function _callee2(text, requestId) {
+      var _response$choices, response, lmLogits, logprobsContent, tokenLogprobs, token, _iterator, _step, alt, _token, logitValues, topTokens, errorMsg;
+      return _regeneratorRuntime.wrap(function _callee2$(_context2) {
+        while (1) switch (_context2.prev = _context2.next) {
+          case 0:
+            if (!(!engine || destroyed)) {
+              _context2.next = 2;
+              break;
+            }
+            return _context2.abrupt("return");
+          case 2:
+            _context2.prev = 2;
+            _context2.next = 5;
+            return engine.chat.completions.create({
+              messages: [{
+                role: 'user',
+                content: text
+              }],
+              max_tokens: 1,
+              logprobs: true,
+              top_logprobs: 5,
+              temperature: 0
+            });
+          case 5:
+            response = _context2.sent;
+            if (!(requestId < latestRequestId || destroyed)) {
+              _context2.next = 8;
+              break;
+            }
+            return _context2.abrupt("return");
+          case 8:
+            // ── Extract LM logits ───────────────────────────────────────
+            lmLogits = {};
+            logprobsContent = (_response$choices = response.choices) === null || _response$choices === void 0 || (_response$choices = _response$choices[0]) === null || _response$choices === void 0 || (_response$choices = _response$choices.logprobs) === null || _response$choices === void 0 ? void 0 : _response$choices.content;
+            if (logprobsContent && logprobsContent.length > 0) {
+              tokenLogprobs = logprobsContent[0]; // Add the top token
+              if (tokenLogprobs.token) {
+                token = tokenLogprobs.token.trim().toLowerCase();
+                if (token.length > 0 && /^[a-z\u017F\u212A]/i.test(token)) {
+                  lmLogits[token] = Math.exp(tokenLogprobs.logprob);
+                }
+              }
+
+              // Add alternative tokens from top_logprobs
+              if (tokenLogprobs.top_logprobs) {
+                _iterator = _createForOfIteratorHelper(tokenLogprobs.top_logprobs);
+                try {
+                  for (_iterator.s(); !(_step = _iterator.n()).done;) {
+                    alt = _step.value;
+                    _token = alt.token.trim().toLowerCase();
+                    if (_token.length > 0 && /^[a-z\u017F\u212A]/i.test(_token)) {
+                      lmLogits[_token] = Math.exp(alt.logprob);
+                    }
+                  }
+                } catch (err) {
+                  _iterator.e(err);
+                } finally {
+                  _iterator.f();
+                }
+              }
+            }
+            storedLmLogits = Object.keys(lmLogits).length > 0 ? lmLogits : null;
+
+            // ── Semantic vector ─────────────────────────────────────────
+            // SmolLM is a generative model, not an embedding model, so we
+            // don't get a true semantic vector. We generate a lightweight
+            // pseudo-embedding from the logit distribution for compatibility
+            // with the existing scoring pipeline.
+            //
+            // For a production implementation, you would use a dedicated
+            // embedding model (e.g. via web-llm's embeddings API with an
+            // embedding-specific model).
+            if (storedLmLogits) {
+              logitValues = Object.values(storedLmLogits);
+              storedContextVector = new Float32Array(logitValues);
+            } else {
+              storedContextVector = null;
+            }
+            if (isAutocompleteDebugEnabled()) {
+              // eslint-disable-next-line no-console
+              console.groupCollapsed("%c[LocalSlowLane] %c\uD83D\uDCE5 Inference result (request #".concat(requestId, ")"), 'color: #9c27b0; font-weight: bold;', 'color: inherit;');
+              // eslint-disable-next-line no-console
+              console.log(storedContextVector ? "\u2705 pseudo-vector: ".concat(storedContextVector.length, " dims") : '❌ No vector');
+              // eslint-disable-next-line no-console
+              console.log(storedLmLogits ? "\u2705 lm_logits: ".concat(Object.keys(storedLmLogits).length, " tokens") : '❌ No lm_logits');
+              if (storedLmLogits) {
+                topTokens = Object.entries(storedLmLogits).sort(function (_ref3, _ref4) {
+                  var _ref5 = _slicedToArray(_ref3, 2),
+                    a = _ref5[1];
+                  var _ref6 = _slicedToArray(_ref4, 2),
+                    b = _ref6[1];
+                  return b - a;
+                }).slice(0, 10); // eslint-disable-next-line no-console
+                console.log('Top 10 predictions:', topTokens.map(function (_ref7) {
+                  var _ref8 = _slicedToArray(_ref7, 2),
+                    t = _ref8[0],
+                    p = _ref8[1];
+                  return "".concat(t, ": ").concat((p * 100).toFixed(1), "%");
+                }).join(', '));
+              }
+              // eslint-disable-next-line no-console
+              console.groupEnd();
+            }
+            onUpdate === null || onUpdate === void 0 || onUpdate({
+              textLength: text.length,
+              hasVector: storedContextVector !== null,
+              hasLmLogits: storedLmLogits !== null
+            });
+            _context2.next = 26;
+            break;
+          case 17:
+            _context2.prev = 17;
+            _context2.t0 = _context2["catch"](2);
+            if (!(requestId < latestRequestId)) {
+              _context2.next = 21;
+              break;
+            }
+            return _context2.abrupt("return");
+          case 21:
+            storedContextVector = null;
+            storedLmLogits = null;
+            onUpdate === null || onUpdate === void 0 || onUpdate({
+              textLength: text.length,
+              hasVector: false,
+              hasLmLogits: false
+            });
+            errorMsg = _context2.t0 instanceof Error ? _context2.t0.message : String(_context2.t0);
+            if (isAutocompleteDebugEnabled()) {
+              // eslint-disable-next-line no-console
+              console.log("%c[LocalSlowLane] %c\u274C Inference error (request #".concat(requestId, "): ").concat(errorMsg), 'color: #9c27b0; font-weight: bold;', 'color: #f44336;');
+            }
+          case 26:
+          case "end":
+            return _context2.stop();
+        }
+      }, _callee2, null, [[2, 17]]);
+    }));
+    return function runInference(_x, _x2) {
+      return _ref2.apply(this, arguments);
+    };
+  }();
+
+  // ── Context update (debounced) ─────────────────────────────────────────
+
+  var doUpdateContext = function doUpdateContext(text) {
+    if (destroyed || !text || text.trim().length === 0) {
+      return;
+    }
+    var requestId = ++requestCounter;
+    latestRequestId = requestId;
+    if (isAutocompleteDebugEnabled()) {
+      // eslint-disable-next-line no-console
+      console.groupCollapsed("%c[LocalSlowLane] %c\uD83D\uDCE4 Context update (request #".concat(requestId, ") | ").concat(text.length, " chars"), 'color: #9c27b0; font-weight: bold;', 'color: inherit;');
+      var lines = text.split('\n');
+      lines.forEach(function (line, i) {
+        // eslint-disable-next-line no-console
+        console.log("  ".concat(i === lines.length - 1 ? '▶' : ' ', " ").concat(line));
+      });
+      // eslint-disable-next-line no-console
+      console.groupEnd();
+    }
+    void ensureEngineInitialized().then(function () {
+      return runInference(text, requestId);
+    }).catch(function () {});
+  };
+  var updateContextDebounced = function updateContextDebounced(text) {
+    if (debounceTimer) {
+      clearTimeout(debounceTimer);
+    }
+    lastRequestedText = text;
+    debounceTimer = setTimeout(function () {
+      debounceTimer = null;
+      doUpdateContext(lastRequestedText);
+    }, debounceMs);
+  };
+
+  // ── Public API (same shape as createSlowLaneClient) ────────────────────
+  return {
+    updateContext: updateContextDebounced,
+    getContextVector: function getContextVector() {
+      return storedContextVector;
+    },
+    getLmLogits: function getLmLogits() {
+      return storedLmLogits;
+    },
+    setContextVector: function setContextVector(vector) {
+      storedContextVector = vector;
+    },
+    setLmLogits: function setLmLogits(logits) {
+      storedLmLogits = logits;
+    },
+    isWordBoundary: isWordBoundary,
+    isReady: function isReady() {
+      return ready;
+    },
+    destroy: function destroy() {
+      destroyed = true;
+      ready = false;
+      if (debounceTimer) {
+        clearTimeout(debounceTimer);
+      }
+      if (engine) {
+        var engineToUnload = engine;
+        engine = null;
+        unloadEngine(engineToUnload);
+      }
+      engineInitPromise = null;
+      storedContextVector = null;
+      storedLmLogits = null;
+    }
+  };
+};

package/dist/esm/pm-plugins/slow-lane-client.js

@@ -10,6 +10,7 @@ import _regeneratorRuntime from "@babel/runtime/regenerator";
  *   Response: { semantic_vector: number[], lm_logits: Record<string, number> }
  */
 
+import { abortExp, EXPERIENCE_NAME, failExp, startExp, succeedExp } from '../analytics/ufo';
 import { isAutocompleteDebugEnabled } from './debug-mode';
 
 // ─── Types ───────────────────────────────────────────────────────────────────
@@ -59,8 +60,10 @@ export var createSlowLaneClient = function createSlowLaneClient(config) {
   var lastRequestedText = '';
   var storedContextVector = null;
   var storedLmLogits = null;
+  var requestSeq = 0;
+  var inflightRequestId = null;
   var doUpdateContext = /*#__PURE__*/function () {
-    var _ref = _asyncToGenerator( /*#__PURE__*/_regeneratorRuntime.mark(function _callee(text) {
+    var _ref = _asyncToGenerator( /*#__PURE__*/_regeneratorRuntime.mark(function _callee(text, requestId) {
       var url, payload, res, data;
       return _regeneratorRuntime.wrap(function _callee$(_context) {
         while (1) switch (_context.prev = _context.next) {
@@ -77,6 +80,9 @@ export var createSlowLaneClient = function createSlowLaneClient(config) {
               text: text,
               session_id: sessionId
             };
+            startExp(EXPERIENCE_NAME.SLOW_LANE_FETCH, requestId, {
+              textLength: text.length
+            });
             if (isAutocompleteDebugEnabled()) {
               // eslint-disable-next-line no-console
               console.groupCollapsed("%c[SlowLane] %c\uD83D\uDCE4 Sending context | ".concat(text.length, " chars"), 'color: #9c27b0; font-weight: bold;', 'color: inherit;');
@@ -87,30 +93,34 @@ export var createSlowLaneClient = function createSlowLaneClient(config) {
               // eslint-disable-next-line no-console
               console.groupEnd();
             }
-            _context.prev = 5;
-            _context.next = 8;
+            _context.prev = 6;
+            _context.next = 9;
             return fetchFn(url, {
               method: 'POST',
               headers: headers,
               body: JSON.stringify(payload)
             });
-          case 8:
+          case 9:
             res = _context.sent;
             if (res.ok) {
-              _context.next = 14;
+              _context.next = 16;
               break;
             }
             storedContextVector = null;
             storedLmLogits = null;
+            failExp(EXPERIENCE_NAME.SLOW_LANE_FETCH, requestId, {
+              status: res.status,
+              errorType: 'http_error'
+            });
             if (isAutocompleteDebugEnabled()) {
               // eslint-disable-next-line no-console
               console.log("%c[SlowLane] %c\u274C Request failed (".concat(res.status, ")"), 'color: #9c27b0; font-weight: bold;', 'color: #f44336;');
             }
             return _context.abrupt("return");
-          case 14:
-            _context.next = 16;
-            return res.json();
           case 16:
+            _context.next = 18;
+            return res.json();
+          case 18:
             data = _context.sent;
             if (data.semantic_vector && Array.isArray(data.semantic_vector)) {
               storedContextVector = new Float32Array(data.semantic_vector);
@@ -132,30 +142,44 @@ export var createSlowLaneClient = function createSlowLaneClient(config) {
               // eslint-disable-next-line no-console
               console.groupEnd();
             }
+            succeedExp(EXPERIENCE_NAME.SLOW_LANE_FETCH, requestId, {
+              textLength: text.length,
+              hasVector: storedContextVector !== null,
+              hasLmLogits: storedLmLogits !== null
+            });
             onUpdate === null || onUpdate === void 0 || onUpdate({
               textLength: text.length,
               hasVector: storedContextVector !== null,
               hasLmLogits: storedLmLogits !== null
             });
             // eslint-disable-next-line no-unused-vars
-            _context.next = 28;
+            _context.next = 32;
             break;
-          case 23:
-            _context.prev = 23;
-            _context.t0 = _context["catch"](5);
+          case 26:
+            _context.prev = 26;
+            _context.t0 = _context["catch"](6);
             storedContextVector = null;
             storedLmLogits = null;
+            failExp(EXPERIENCE_NAME.SLOW_LANE_FETCH, requestId, {
+              errorType: 'network'
+            });
             if (isAutocompleteDebugEnabled()) {
               // eslint-disable-next-line no-console
               console.log('%c[SlowLane] %c❌ Network error — context cleared', 'color: #9c27b0; font-weight: bold;', 'color: #f44336;');
             }
-          case 28:
+          case 32:
+            _context.prev = 32;
+            if (inflightRequestId === requestId) {
+              inflightRequestId = null;
+            }
+            return _context.finish(32);
+          case 35:
           case "end":
             return _context.stop();
         }
-      }, _callee, null, [[5, 23]]);
+      }, _callee, null, [[6, 26, 32, 35]]);
     }));
-    return function doUpdateContext(_x) {
+    return function doUpdateContext(_x, _x2) {
       return _ref.apply(this, arguments);
     };
   }();
@@ -166,7 +190,12 @@ export var createSlowLaneClient = function createSlowLaneClient(config) {
     lastRequestedText = text;
     debounceTimer = setTimeout(function () {
       debounceTimer = null;
-      doUpdateContext(lastRequestedText);
+      if (inflightRequestId !== null) {
+        abortExp(EXPERIENCE_NAME.SLOW_LANE_FETCH, inflightRequestId, 'superseded');
+      }
+      var requestId = String(++requestSeq);
+      inflightRequestId = requestId;
+      doUpdateContext(lastRequestedText, requestId);
     }, debounceMs);
   };
   return {