@empiricalrun/llm - npm Package Compare versions

Comparing version 0.9.18 to 0.9.19

CHANGELOG.md

 # @empiricalrun/llm
 
+## 0.9.19
+
+### Patch Changes
+
+- 2dafa69: feat: added tracing in llm package
+
 ## 0.9.18
@@ -4,0 +10,0 @@

dist/trace/index.d.ts

 import { Langfuse } from "langfuse";
+export declare const isLangfuseEnabled: string | undefined;
 export declare const langfuseInstance: Langfuse;
 /**
+ * This is useful when you want to use langfuse only if it is enabled otherwise silently fail.
+ */
+export declare function getLangfuseInstanceIfEnabled(): Langfuse | undefined;
+/**
  * Flush all traces to trace service.
@@ -5,0 +10,0 @@ * This is useful for debugging purposes.

dist/trace/index.js

 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.flushAllTraces = exports.langfuseInstance = void 0;
+exports.flushAllTraces = exports.getLangfuseInstanceIfEnabled = exports.langfuseInstance = exports.isLangfuseEnabled = void 0;
 const langfuse_1 = require("langfuse");
+exports.isLangfuseEnabled = process.env.LANGFUSE_SK && process.env.LANGFUSE_PK;
 exports.langfuseInstance = new langfuse_1.Langfuse({
@@ -12,2 +13,9 @@ secretKey: process.env.LANGFUSE_SK,
 /**
+ * This is useful when you want to use langfuse only if it is enabled otherwise silently fail.
+ */
+function getLangfuseInstanceIfEnabled() {
+    return exports.isLangfuseEnabled ? exports.langfuseInstance : undefined;
+}
+exports.getLangfuseInstanceIfEnabled = getLangfuseInstanceIfEnabled;
+/**
  * Flush all traces to trace service.
@@ -22,4 +30,4 @@ * This is useful for debugging purposes.
 async function flushAllTraces() {
-    await exports.langfuseInstance.flushAsync();
+    await exports.langfuseInstance?.flushAsync();
 }
 exports.flushAllTraces = flushAllTraces;

dist/vision/bbox/index.d.ts

 import { BoundingBox } from "../types";
 export declare function getBoundingBox(base64Image: string, elementDescription: string, options?: {
     debug?: true;
+    telemetry?: {
+        tags?: string[];
+    };
 }): Promise<BoundingBox>;
 //# sourceMappingURL=index.d.ts.map

dist/vision/bbox/index.js

@@ -5,2 +5,3 @@ "use strict";
 const __1 = require("../..");
+const trace_1 = require("../../trace");
 const image_1 = require("../image");
@@ -16,3 +17,35 @@ const utils_1 = require("../utils");
     });
-    const content = `You are given a screenshot and a description of an element on the screen. You are an expert vision model that can identify bounding box for the element on the screen. Return the output as an array [y_min, x_min, y_max, x_max]. Return only the array and nothing else.\n\nElement description:`;
+    const langfuseInstance = (0, trace_1.getLangfuseInstanceIfEnabled)();
+    if (options?.telemetry?.tags && !langfuseInstance) {
+        console.warn("Telemetry keys are not set. Telemetry will not be sent.");
+    }
+    const session = (0, utils_1.getSessionDetails)();
+    const trace = langfuseInstance?.trace({
+        name: "vision-bbox",
+        sessionId: session.id,
+        release: session.version,
+        tags: ["vision", "vision-bbox", ...(options?.telemetry?.tags ?? [])],
+    });
+    trace?.update({
+        input: {
+            base64Image,
+            elementDescription,
+        },
+    });
+    const temperature = 0.5;
+    const model = "gemini-1.5-pro-latest";
+    const content = `You are given a screenshot and a description of an element on the screen. You are an expert vision model that can identify bounding box for the element on the screen. Return the output as an array [y_min, x_min, y_max, x_max]. Return only the array and nothing else.\n\nElement description:\n${elementDescription}`;
+    const generationTrace = trace?.generation({
+        name: "get-bounding-box",
+        model,
+        modelParameters: {
+            temperature,
+        },
+        input: {
+            systemPrompt: content,
+            userMessages: {
+                image: base64Image,
+            },
+        },
+    });
     const llmResponse = await llm.createChatCompletion({
@@ -22,3 +55,3 @@ messages: [
                 role: "system",
-                content: `${content}\n${elementDescription}`,
+                content,
             },
@@ -37,9 +70,17 @@ {
         ],
-        model: "gemini-1.5-pro-latest",
+        model,
         modelParameters: {
-            temperature: 0.5,
+            temperature,
         },
     });
+    trace?.update({
+        output: llmResponse?.content,
+    });
+    generationTrace?.end({
+        output: {
+            queryResponse: llmResponse?.content,
+        },
+    });
     if (!llmResponse) {
-        throw new Error("Failed to get bounding box");
+        throw new Error(`Failed to get bounding box. Check the trace for more info: ${trace?.getTraceUrl()}`);
     }
@@ -49,3 +90,3 @@ const message = llmResponse.content;
     if (!arrayAsString) {
-        throw new Error(`Failed to find array in LLM response: ${message}`);
+        throw new Error(`Failed to find array in LLM response: ${message}. Check the trace for more info: ${trace?.getTraceUrl()}`);
     }
@@ -58,3 +99,3 @@ let bbox;
     catch (e) {
-        throw new Error(`Failed to parse LLM response: ${message}`);
+        throw new Error(`Failed to parse LLM response: ${message}. Check the trace for more info: ${trace?.getTraceUrl()}`);
     }
@@ -61,0 +102,0 @@ if (options.debug) {

dist/vision/point/index.d.ts

@@ -0,1 +1,2 @@
+import { TraceClient } from "../..";
 import { Coordinates, Point } from "../types";
@@ -9,5 +10,8 @@ /**
     useCache?: boolean;
+    telemetry?: {
+        tags?: string[];
+    };
 }): Promise<Coordinates>;
-export declare function extractTapCoordinateFromString(inputString: string): Point;
+export declare function extractTapCoordinateFromString(inputString: string, trace?: TraceClient): Point;
 export declare function getLlmResponse(base64Image: string, prompt: string): Promise<string>;
 //# sourceMappingURL=index.d.ts.map

dist/vision/point/index.js

@@ -8,3 +8,5 @@ "use strict";
 const async_retry_1 = __importDefault(require("async-retry"));
+const trace_1 = require("../../trace");
 const image_1 = require("../image");
+const utils_1 = require("../utils");
 const cache_1 = require("./cache");
@@ -28,4 +30,31 @@ const API_BASE_URL = process.env.VISION_MODEL_ENDPOINT;
     if (!scaledPoint) {
+        const langfuseInstance = (0, trace_1.getLangfuseInstanceIfEnabled)();
+        if (options?.telemetry?.tags && !langfuseInstance) {
+            console.warn("Telemetry keys are not set. Telemetry will not be sent.");
+        }
+        const session = (0, utils_1.getSessionDetails)();
+        const trace = langfuseInstance?.trace({
+            name: "vision-point",
+            sessionId: session.id,
+            release: session.version,
+        });
+        trace?.update({
+            tags: ["vision", "vision-point", ...(options?.telemetry?.tags ?? [])],
+            input: {
+                base64Image,
+                prompt,
+            },
+        });
+        const generationTrace = trace?.generation({
+            name: "get-coordinates-for",
+            model: "Molmo-7B-D-0924",
+            input: {
+                userMessages: {
+                    prompt,
+                    image: base64Image,
+                },
+            },
+        });
         const llmResponse = await getLlmResponse(base64Image, prompt);
-        const pointFromLlm = extractTapCoordinateFromString(llmResponse);
+        const pointFromLlm = extractTapCoordinateFromString(llmResponse, trace);
         scaledPoint = scaleForImage(pointFromLlm, base64Image);
@@ -35,2 +64,8 @@ if (options?.useCache) {
         }
+        trace?.update({
+            output: scaledPoint,
+        });
+        generationTrace?.end({
+            output: scaledPoint,
+        });
     }
@@ -64,3 +99,3 @@ const annotatedImage = await (0, image_1.drawRedDotAtPoint)(base64Image, scaledPoint);
 }
-function extractTapCoordinateFromString(inputString) {
+function extractTapCoordinateFromString(inputString, trace) {
     inputString = inputString.trim();
@@ -86,3 +121,3 @@ if (inputString.startsWith('"') && inputString.endsWith('"')) {
     }
-    throw new Error(`No valid coordinates found in the input string: ${inputString}`);
+    throw new Error(`No valid coordinates found in the input string: ${inputString}. ${trace ? `Check the trace for more info: ${trace.getTraceUrl()}` : ""}`);
 }
@@ -89,0 +124,0 @@ exports.extractTapCoordinateFromString = extractTapCoordinateFromString;

dist/vision/query/index.d.ts

@@ -7,4 +7,7 @@ import { z, ZodType } from "zod";
     model?: LLMModel;
+    telemetry?: {
+        tags?: string[];
+    };
 }): Promise<ExtractType<T>>;
 export {};
 //# sourceMappingURL=index.d.ts.map

dist/vision/query/index.js

 "use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
 Object.defineProperty(exports, "__esModule", { value: true });
@@ -6,3 +9,5 @@ exports.query = void 0;
 const zod_2 = require("zod");
+const zod_to_json_schema_1 = __importDefault(require("zod-to-json-schema"));
 const __1 = require("../..");
+const trace_1 = require("../../trace");
 const utils_1 = require("../utils");
@@ -13,3 +18,8 @@ async function query(base64Image, instruction, options = {
 }) {
-    const { responseFormat = zod_2.z.string(), model = "gpt-4o-mini" } = options;
+    const mergedOptions = {
+        responseFormat: zod_2.z.string(),
+        model: "gpt-4o-mini",
+        ...options,
+    };
+    const { responseFormat, model } = mergedOptions;
     const llm = new __1.LLM({
@@ -35,2 +45,39 @@ provider: "openai",
     }
+    const session = (0, utils_1.getSessionDetails)();
+    const langfuseInstance = (0, trace_1.getLangfuseInstanceIfEnabled)();
+    if (mergedOptions.telemetry?.tags && !langfuseInstance) {
+        console.warn("Telemetry keys are not set. Telemetry will not be sent.");
+    }
+    const trace = langfuseInstance?.trace({
+        name: "vision-query",
+        sessionId: session.id,
+        release: session.version,
+        tags: ["vision", "vision-query", ...(mergedOptions?.telemetry?.tags || [])],
+    });
+    const temperature = 0.5;
+    const generationTrace = trace?.generation({
+        name: "query",
+        model,
+        modelParameters: {
+            temperature,
+        },
+        input: {
+            responseFormat: (0, zod_to_json_schema_1.default)(extendedResponseFormat, "responseFormat"),
+            systemPrompt,
+            userMessages: {
+                instruction,
+                image: base64Image,
+            },
+        },
+    });
+    trace?.update({
+        input: {
+            base64Image,
+            instruction,
+            options: {
+                responseFormat: (0, zod_to_json_schema_1.default)(extendedResponseFormat, "responseFormat"),
+                model,
+            },
+        },
+    });
     const llmResponse = await llm.createChatCompletion({
@@ -60,3 +107,3 @@ messages: [
         modelParameters: {
-            temperature: 0.5,
+            temperature: temperature,
         },
@@ -66,8 +113,19 @@ responseFormat: (0, zod_1.zodResponseFormat)(extendedResponseFormat, "your_response"),
     if (!llmResponse || !llmResponse.content) {
-        throw new Error("Query failed: no response content from LLM");
+        throw new Error(`Query failed: no response content from LLM. Check the trace for more info: ${trace?.getTraceUrl()}`);
     }
     const response = llmResponse.content;
     const jsonData = JSON.parse(response);
+    trace?.update({
+        output: jsonData,
+    });
+    generationTrace?.end({
+        output: jsonData,
+        usage: {
+            input: llm.promptTokens,
+            output: llm.completionTokens,
+            unit: "TOKENS",
+        },
+    });
     return responseFormat.parse(jsonData.answer);
 }
 exports.query = query;

dist/vision/utils.d.ts

 import { LLMProvider } from "..";
 export declare function imageFormatForProvider(provider: LLMProvider, image: string): string;
+export declare const getSessionDetails: () => {
+    id: `${string}-${string}-${string}-${string}-${string}`;
+    version: string;
+};
 //# sourceMappingURL=utils.d.ts.map

dist/vision/utils.js

 "use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.imageFormatForProvider = void 0;
+exports.getSessionDetails = exports.imageFormatForProvider = void 0;
+const package_json_1 = __importDefault(require("../../package.json"));
 function imageFormatForProvider(provider, image) {
@@ -14,1 +18,9 @@ let base64Image = image;
 exports.imageFormatForProvider = imageFormatForProvider;
+const getSessionDetails = () => {
+    const sessionId = crypto.randomUUID();
+    return {
+        id: sessionId,
+        version: package_json_1.default.version,
+    };
+};
+exports.getSessionDetails = getSessionDetails;

package.json

 {
   "name": "@empiricalrun/llm",
-  "version": "0.9.18",
+  "version": "0.9.19",
   "main": "dist/index.js",
@@ -39,3 +39,4 @@ "exports": {
     "sharp": "^0.33.5",
-    "zod": "^3.23.8"
+    "zod": "^3.23.8",
+    "zod-to-json-schema": "^3.23.5"
   },
@@ -42,0 +43,0 @@ "devDependencies": {

No alert changes

Improved metrics

Total package byte prevSize

48759

increased by15.38%

Lines of code

913

increased by23.38%

Worsened metrics

Dependency count

8

increased by14.29%

Number of low supply chain risk alerts

9

increased by28.57%

Dependency changes

• + Addedzod-to-json-schema@^3.23.5

• + Addedzod-to-json-schema@3.24.1(transitive)