diff --git a/.changeset/rich-coins-leave.md b/.changeset/rich-coins-leave.md
new file mode 100644
index 0000000000000000000000000000000000000000..44da9104700ff11bc820409be26b4581a0f37e87
--- /dev/null
+++ b/.changeset/rich-coins-leave.md
@@ -0,0 +1,5 @@
+---
+"@llamaindex/core": minor
+---
+
+add defaultQuestionExtractPrompt
diff --git a/.changeset/three-gorillas-do.md b/.changeset/three-gorillas-do.md
new file mode 100644
index 0000000000000000000000000000000000000000..72a69c0c5de2c560e22e03c60d9c4665b746f8a4
--- /dev/null
+++ b/.changeset/three-gorillas-do.md
@@ -0,0 +1,5 @@
+---
+"llamaindex": minor
+---
+
+Correct initialization of QuestionsAnsweredExtractor so that it uses the promptTemplate arg when passed in
diff --git a/packages/core/src/prompts/index.ts b/packages/core/src/prompts/index.ts
index c8ae7087f48b1397d7b816eb4f595019bf51b5de..992411a5b9a8be89b2562da4503aa92f6cd637cb 100644
--- a/packages/core/src/prompts/index.ts
+++ b/packages/core/src/prompts/index.ts
@@ -13,6 +13,7 @@ export {
defaultContextSystemPrompt,
defaultKeywordExtractPrompt,
defaultQueryKeywordExtractPrompt,
+ defaultQuestionExtractPrompt,
defaultRefinePrompt,
defaultSubQuestionPrompt,
defaultSummaryPrompt,
@@ -25,6 +26,7 @@ export type {
ContextSystemPrompt,
KeywordExtractPrompt,
QueryKeywordExtractPrompt,
+ QuestionExtractPrompt,
RefinePrompt,
SubQuestionPrompt,
SummaryPrompt,
diff --git a/packages/core/src/prompts/prompt.ts b/packages/core/src/prompts/prompt.ts
index aeb6a7e642ff82fb872d1e5402483df0f6a73ad9..1efa0289f12a96fb9734fca71d5b6981bf9d663d 100644
--- a/packages/core/src/prompts/prompt.ts
+++ b/packages/core/src/prompts/prompt.ts
@@ -15,6 +15,7 @@ export type CondenseQuestionPrompt = PromptTemplate<
export type ContextSystemPrompt = PromptTemplate<["context"]>;
export type KeywordExtractPrompt = PromptTemplate<["context"]>;
export type QueryKeywordExtractPrompt = PromptTemplate<["question"]>;
+export type QuestionExtractPrompt = PromptTemplate<["context", "numQuestions"]>;
export const defaultTextQAPrompt: TextQAPrompt = new PromptTemplate({
templateVars: ["context", "query"],
@@ -253,3 +254,17 @@ export const defaultQueryKeywordExtractPrompt = new PromptTemplate({
}).partialFormat({
maxKeywords: "10",
});
+
+export const defaultQuestionExtractPrompt = new PromptTemplate({
+ templateVars: ["numQuestions", "context"],
+ template: `(
+ "Given the contextual informations below, generate {numQuestions} questions this context can provides specific answers to which are unlikely to be found else where. Higher-level summaries of surrounding context may be provided as well. "
+ "Try using these summaries to generate better questions that this context can answer."
+ "---------------------"
+ "{context}"
+ "---------------------"
+ "Provide questions in the following format: 'QUESTIONS: <questions>'"
+)`,
+}).partialFormat({
+ numQuestions: "5",
+});
diff --git a/packages/llamaindex/src/extractors/MetadataExtractors.ts b/packages/llamaindex/src/extractors/MetadataExtractors.ts
index b1d0ce5a47412627745c0375ef86b1a0f30e27b9..e43ef941521cfda2602cde871639bc66cbc15a5b 100644
--- a/packages/llamaindex/src/extractors/MetadataExtractors.ts
+++ b/packages/llamaindex/src/extractors/MetadataExtractors.ts
@@ -1,10 +1,14 @@
import type { LLM } from "@llamaindex/core/llms";
+import {
+ PromptTemplate,
+ defaultQuestionExtractPrompt,
+ type QuestionExtractPrompt,
+} from "@llamaindex/core/prompts";
import type { BaseNode } from "@llamaindex/core/schema";
import { MetadataMode, TextNode } from "@llamaindex/core/schema";
import { OpenAI } from "@llamaindex/openai";
import {
defaultKeywordExtractorPromptTemplate,
- defaultQuestionAnswerPromptTemplate,
defaultSummaryExtractorPromptTemplate,
defaultTitleCombinePromptTemplate,
defaultTitleExtractorPromptTemplate,
@@ -247,7 +251,7 @@ export class TitleExtractor extends BaseExtractor {
type QuestionAnswerExtractArgs = {
llm?: LLM;
questions?: number;
- promptTemplate?: string;
+ promptTemplate?: QuestionExtractPrompt["template"];
embeddingOnly?: boolean;
};
@@ -276,7 +280,7 @@ export class QuestionsAnsweredExtractor extends BaseExtractor {
* The prompt template to use for the question extractor.
* @type {string}
*/
- promptTemplate: string;
+ promptTemplate: QuestionExtractPrompt;
/**
* Wheter to use metadata for embeddings only
@@ -289,7 +293,7 @@ export class QuestionsAnsweredExtractor extends BaseExtractor {
* Constructor for the QuestionsAnsweredExtractor class.
* @param {LLM} llm LLM instance.
* @param {number} questions Number of questions to generate.
- * @param {string} promptTemplate The prompt template to use for the question extractor.
+ * @param {TextQAPrompt} promptTemplate The prompt template to use for the question extractor.
* @param {boolean} embeddingOnly Wheter to use metadata for embeddings only.
*/
constructor(options?: QuestionAnswerExtractArgs) {
@@ -300,12 +304,14 @@ export class QuestionsAnsweredExtractor extends BaseExtractor {
this.llm = options?.llm ?? new OpenAI();
this.questions = options?.questions ?? 5;
- this.promptTemplate =
- options?.promptTemplate ??
- defaultQuestionAnswerPromptTemplate({
- numQuestions: this.questions,
- contextStr: "",
- });
+ this.promptTemplate = options?.promptTemplate
+ ? new PromptTemplate({
+ templateVars: ["numQuestions", "context"],
+ template: options.promptTemplate,
+ }).partialFormat({
+ numQuestions: "5",
+ })
+ : defaultQuestionExtractPrompt;
this.embeddingOnly = options?.embeddingOnly ?? false;
}
@@ -323,9 +329,9 @@ export class QuestionsAnsweredExtractor extends BaseExtractor {
const contextStr = node.getContent(this.metadataMode);
- const prompt = defaultQuestionAnswerPromptTemplate({
- contextStr,
- numQuestions: this.questions,
+ const prompt = this.promptTemplate.format({
+ context: contextStr,
+ numQuestions: this.questions.toString(),
});
const questions = await this.llm.complete({
diff --git a/packages/llamaindex/src/extractors/prompts.ts b/packages/llamaindex/src/extractors/prompts.ts
index 5c2d5448ac5e66f4577cdafa0bcca90ca71705d7..0065e5d834887588caa3da1c80f6e5634285d48c 100644
--- a/packages/llamaindex/src/extractors/prompts.ts
+++ b/packages/llamaindex/src/extractors/prompts.ts
@@ -7,11 +7,6 @@ export interface DefaultKeywordExtractorPromptTemplate
keywords: number;
}
-export interface DefaultQuestionAnswerPromptTemplate
- extends DefaultPromptTemplate {
- numQuestions: number;
-}
-
export interface DefaultNodeTextTemplate {
metadataStr: string;
content: string;
@@ -41,16 +36,6 @@ export const defaultTitleCombinePromptTemplate = (
Based on the above candidate titles and contents, what is the comprehensive title for this document?
Title: `;
-export const defaultQuestionAnswerPromptTemplate = (
- { contextStr = "", numQuestions = 5 }: DefaultQuestionAnswerPromptTemplate = {
- contextStr: "",
- numQuestions: 5,
- },
-) => `${contextStr}
-Given the contextual informations, generate ${numQuestions} questions this context can provides specific answers to which are unlikely to be found else where. Higher-level summaries of surrounding context may be provideds as well.
-Try using these summaries to generate better questions that this context can answer.
-`;
-
export const defaultSummaryExtractorPromptTemplate = (
{ contextStr = "" }: DefaultPromptTemplate = {
contextStr: "",
diff --git a/packages/llamaindex/tests/MetadataExtractors.test.ts b/packages/llamaindex/tests/MetadataExtractors.test.ts
index fe0e76c95607cddef1cd9c22c64abcd12946baf5..6cda9693866157d6aad6265cfa8b3df1c2da1b9e 100644
--- a/packages/llamaindex/tests/MetadataExtractors.test.ts
+++ b/packages/llamaindex/tests/MetadataExtractors.test.ts
@@ -102,6 +102,35 @@ describe("[MetadataExtractor]: Extractors should populate the metadata", () => {
});
});
+ test("[MetadataExtractor] QuestionsAnsweredExtractor uses custom prompt template", async () => {
+ const nodeParser = new SentenceSplitter();
+
+ const nodes = nodeParser.getNodesFromDocuments([
+ new Document({ text: DEFAULT_LLM_TEXT_OUTPUT }),
+ ]);
+
+ const llmCompleteSpy = vi.spyOn(serviceContext.llm, "complete");
+
+ const questionsAnsweredExtractor = new QuestionsAnsweredExtractor({
+ llm: serviceContext.llm,
+ questions: 5,
+ promptTemplate: `This is a custom prompt template for {context} with {numQuestions} questions`,
+ });
+
+ await questionsAnsweredExtractor.processNodes(nodes);
+
+ expect(llmCompleteSpy).toHaveBeenCalled();
+
+ // Build the expected prompt
+ const expectedPrompt = `This is a custom prompt template for ${DEFAULT_LLM_TEXT_OUTPUT} with 5 questions`;
+
+ // Get the actual prompt used in llm.complete
+ const actualPrompt = llmCompleteSpy.mock?.calls?.[0]?.[0];
+
+ // Assert that the prompts match
+ expect(actualPrompt).toEqual({ prompt: expectedPrompt });
+ });
+
test("[MetadataExtractor] SumamryExtractor returns sectionSummary metadata", async () => {
const nodeParser = new SentenceSplitter();