@promptbook/node 0.92.0-5 → 0.92.0-8

package/umd/index.umd.js

@@ -46,7 +46,7 @@
      * @generated
      * @see https://github.com/webgptorg/promptbook
      */
-    const PROMPTBOOK_ENGINE_VERSION = '0.92.0-5';
+    const PROMPTBOOK_ENGINE_VERSION = '0.92.0-8';
     /**
      * TODO: string_promptbook_version should be constrained to the all versions of Promptbook engine
      * Note: [💞] Ignore a discrepancy between file name and entity name
@@ -315,6 +315,45 @@
         }
     }
 
+    /**
+     * 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__default["default"](`
+                    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__default["default"]((block) => `
+                    ${block(error.message)}
+            
+                    The JSON text:
+                    ${block(value)}
+                `));
+        }
+    }
+    /**
+     * TODO: !!!! Use in Promptbook.studio
+     */
+
     /**
      * Orders JSON object by keys
      *
@@ -1104,7 +1143,7 @@
         if (!indexFile) {
             throw new UnexpectedError(`Archive does not contain 'index.book.json' file`);
         }
-        const collectionJson = JSON.parse(await indexFile.async('text'));
+        const collectionJson = jsonParse(await indexFile.async('text'));
         for (const pipeline of collectionJson) {
             validatePipeline(pipeline);
         }
@@ -1698,7 +1737,7 @@
         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);
@@ -4092,13 +4131,79 @@
     /**
      * @@@
      *
+     * 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
+     */
 
     /**
      * @@@
@@ -4106,9 +4211,9 @@
      * @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;
@@ -4170,6 +4275,7 @@
         }
         const definedParameters = Object.freeze({
             ...(await getReservedParametersForTask({
+                tools,
                 preparedPipeline,
                 task: currentTask,
                 pipelineIdentification,
@@ -4741,18 +4847,26 @@
         }).asPromise();
         const { outputParameters } = result;
         const { modelsRequirements: modelsRequirementsJson } = outputParameters;
-        const modelsRequirementsUnchecked = JSON.parse(modelsRequirementsJson);
+        let modelsRequirementsUnchecked = jsonParse(modelsRequirementsJson);
         if (isVerbose) {
             console.info(`PERSONA ${personaDescription}`, modelsRequirementsUnchecked);
         }
         if (!Array.isArray(modelsRequirementsUnchecked)) {
-            throw new UnexpectedError(spaceTrim__default["default"]((block) => `
-                    Invalid \`modelsRequirements\`:
+            // <- TODO: Book should have syntax and system to enforce shape of JSON
+            modelsRequirementsUnchecked = [modelsRequirementsUnchecked];
+            /*
+            throw new UnexpectedError(
+                spaceTrim(
+                    (block) => `
+                        Invalid \`modelsRequirements\`:
 
-                    \`\`\`json
-                    ${block(JSON.stringify(modelsRequirementsUnchecked, null, 4))}
-                    \`\`\`
-                `));
+                        \`\`\`json
+                        ${block(JSON.stringify(modelsRequirementsUnchecked, null, 4))}
+                        \`\`\`
+                    `,
+                ),
+            );
+            */
         }
         const modelsRequirements = modelsRequirementsUnchecked.map((modelRequirements) => ({
             modelVariant: 'CHAT',
@@ -5322,7 +5436,7 @@
                     > },
                 */
                 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');
@@ -10842,7 +10956,7 @@
                 return null;
             }
             const fileContent = await promises.readFile(filename, 'utf-8');
-            const value = JSON.parse(fileContent);
+            const value = jsonParse(fileContent);
             // TODO: [🌗]
             return value;
         }