@promptbook/pdf 0.92.0-5 → 0.92.0-6

package/esm/index.es.js

@@ -26,7 +26,7 @@ const BOOK_LANGUAGE_VERSION = '1.0.0';
  * @generated
  * @see https://github.com/webgptorg/promptbook
  */
-const PROMPTBOOK_ENGINE_VERSION = '0.92.0-5';
+const PROMPTBOOK_ENGINE_VERSION = '0.92.0-6';
 /**
  * TODO: string_promptbook_version should be constrained to the all versions of Promptbook engine
  * Note: [💞] Ignore a discrepancy between file name and entity name
@@ -2137,6 +2137,45 @@ function isPipelinePrepared(pipeline) {
  *     - [♨] Are tasks prepared
  */
 
+/**
+ * Converts a JavaScript Object Notation (JSON) string into an object.
+ *
+ * Note: This is wrapper around `JSON.parse()` with better error and type handling
+ *
+ * @public exported from `@promptbook/utils`
+ */
+function jsonParse(value) {
+    if (value === undefined) {
+        throw new Error(`Can not parse JSON from undefined value.`);
+    }
+    else if (typeof value !== 'string') {
+        console.error('Can not parse JSON from non-string value.', { text: value });
+        throw new Error(spaceTrim(`
+                    Can not parse JSON from non-string value.
+
+                    The value type: ${typeof value}
+                    See more in console.
+                `));
+    }
+    try {
+        return JSON.parse(value);
+    }
+    catch (error) {
+        if (!(error instanceof Error)) {
+            throw error;
+        }
+        throw new Error(spaceTrim((block) => `
+                    ${block(error.message)}
+            
+                    The JSON text:
+                    ${block(value)}
+                `));
+    }
+}
+/**
+ * TODO: !!!! Use in Promptbook.studio
+ */
+
 /**
  * Recursively converts JSON strings to JSON objects
 
@@ -2155,7 +2194,7 @@ function jsonStringsToJsons(object) {
     const newObject = { ...object };
     for (const [key, value] of Object.entries(object)) {
         if (typeof value === 'string' && isValidJsonString(value)) {
-            newObject[key] = JSON.parse(value);
+            newObject[key] = jsonParse(value);
         }
         else {
             newObject[key] = jsonStringsToJsons(value);
@@ -3002,7 +3041,7 @@ async function preparePersona(personaDescription, tools, options) {
     }).asPromise();
     const { outputParameters } = result;
     const { modelsRequirements: modelsRequirementsJson } = outputParameters;
-    const modelsRequirementsUnchecked = JSON.parse(modelsRequirementsJson);
+    const modelsRequirementsUnchecked = jsonParse(modelsRequirementsJson);
     if (isVerbose) {
         console.info(`PERSONA ${personaDescription}`, modelsRequirementsUnchecked);
     }
@@ -3458,7 +3497,7 @@ async function makeKnowledgeSourceHandler(knowledgeSource, tools, options) {
                 > },
             */
             async asJson() {
-                return JSON.parse(await tools.fs.readFile(filename, 'utf-8'));
+                return jsonParse(await tools.fs.readFile(filename, 'utf-8'));
             },
             async asText() {
                 return await tools.fs.readFile(filename, 'utf-8');
@@ -5145,13 +5184,79 @@ async function getExamplesForTask(task) {
 /**
  * @@@
  *
+ * Here is the place where RAG (retrieval-augmented generation) happens
+ *
  * @private internal utility of `createPipelineExecutor`
  */
 async function getKnowledgeForTask(options) {
-    const { preparedPipeline, task } = options;
-    return preparedPipeline.knowledgePieces.map(({ content }) => `- ${content}`).join('\n');
+    const { tools, preparedPipeline, task } = options;
+    const firstKnowlegePiece = preparedPipeline.knowledgePieces[0];
+    const firstKnowlegeIndex = firstKnowlegePiece === null || firstKnowlegePiece === void 0 ? void 0 : firstKnowlegePiece.index[0];
+    // <- TODO: Do not use just first knowledge piece and first index to determine embedding model, use also keyword search
+    if (firstKnowlegePiece === undefined || firstKnowlegeIndex === undefined) {
+        return 'No knowledge pieces found';
+    }
+    // TODO: [🚐] Make arrayable LLMs -> single LLM DRY
+    const _llms = arrayableToArray(tools.llm);
+    const llmTools = _llms.length === 1 ? _llms[0] : joinLlmExecutionTools(..._llms);
+    const taskEmbeddingPrompt = {
+        title: 'Knowledge Search',
+        modelRequirements: {
+            modelVariant: 'EMBEDDING',
+            modelName: firstKnowlegeIndex.modelName,
+        },
+        content: task.content,
+        parameters: {
+        /* !!!!!!!! */
+        },
+    };
+    const taskEmbeddingResult = await llmTools.callEmbeddingModel(taskEmbeddingPrompt);
+    const knowledgePiecesWithRelevance = preparedPipeline.knowledgePieces.map((knowledgePiece) => {
+        const { index } = knowledgePiece;
+        const knowledgePieceIndex = index.find((i) => i.modelName === firstKnowlegeIndex.modelName);
+        // <- TODO: Do not use just first knowledge piece and first index to determine embedding model
+        if (knowledgePieceIndex === undefined) {
+            return {
+                content: knowledgePiece.content,
+                relevance: 0,
+            };
+        }
+        const relevance = computeCosineSimilarity(knowledgePieceIndex.position, taskEmbeddingResult.content);
+        return {
+            content: knowledgePiece.content,
+            relevance,
+        };
+    });
+    const knowledgePiecesSorted = knowledgePiecesWithRelevance.sort((a, b) => a.relevance - b.relevance);
+    const knowledgePiecesLimited = knowledgePiecesSorted.slice(0, 5);
+    console.log('!!! Embedding', {
+        task,
+        taskEmbeddingPrompt,
+        taskEmbeddingResult,
+        firstKnowlegePiece,
+        firstKnowlegeIndex,
+        knowledgePiecesWithRelevance,
+        knowledgePiecesSorted,
+        knowledgePiecesLimited,
+    });
+    return knowledgePiecesLimited.map(({ content }) => `- ${content}`).join('\n');
     //                                                      <- TODO: [🧠] Some smart aggregation of knowledge pieces, single-line vs multi-line vs mixed
 }
+// TODO: !!!!!! Annotate + to new file
+function computeCosineSimilarity(embeddingVector1, embeddingVector2) {
+    if (embeddingVector1.length !== embeddingVector2.length) {
+        throw new TypeError('Embedding vectors must have the same length');
+    }
+    const dotProduct = embeddingVector1.reduce((sum, value, index) => sum + value * embeddingVector2[index], 0);
+    const magnitude1 = Math.sqrt(embeddingVector1.reduce((sum, value) => sum + value * value, 0));
+    const magnitude2 = Math.sqrt(embeddingVector2.reduce((sum, value) => sum + value * value, 0));
+    return 1 - dotProduct / (magnitude1 * magnitude2);
+}
+/**
+ * TODO: !!!! Verify if this is working
+ * TODO: [♨] Implement Better - use keyword search
+ * TODO: [♨] Examples of values
+ */
 
 /**
  * @@@
@@ -5159,9 +5264,9 @@ async function getKnowledgeForTask(options) {
  * @private internal utility of `createPipelineExecutor`
  */
 async function getReservedParametersForTask(options) {
-    const { preparedPipeline, task, pipelineIdentification } = options;
+    const { tools, preparedPipeline, task, pipelineIdentification } = options;
     const context = await getContextForTask(); // <- [🏍]
-    const knowledge = await getKnowledgeForTask({ preparedPipeline, task });
+    const knowledge = await getKnowledgeForTask({ tools, preparedPipeline, task });
     const examples = await getExamplesForTask();
     const currentDate = new Date().toISOString(); // <- TODO: [🧠][💩] Better
     const modelName = RESERVED_PARAMETER_MISSING_VALUE;
@@ -5223,6 +5328,7 @@ async function executeTask(options) {
     }
     const definedParameters = Object.freeze({
         ...(await getReservedParametersForTask({
+            tools,
             preparedPipeline,
             task: currentTask,
             pipelineIdentification,