fix(ai-bot): better prompt template typing

Author: ascariandrea

packages/@liexp/backend/src/flows/ai/createEventFromDocuments.flow.ts

@@ -22,7 +22,12 @@ export const createEventFromDocuments = <
 >(
   content: Document[],
   type: EventType,
-  prompt: PromptFn,
+  prompt: PromptFn<{
+    type: EventType;
+    jsonSchema: string;
+    question: string;
+    context: string;
+  }>,
   jsonSchema: unknown,
   question: string | null,
 ): ReaderTaskEither<C, APIError, Event> => {

packages/@liexp/backend/src/flows/ai/createEventFromText.flow.ts

@@ -19,14 +19,26 @@ import { runRagChain } from "./runRagChain.js";
 
 export const getCreateEventPromptPartial =
   <C extends LoggerContext>(
-    promptTemplate: PromptFn,
+    promptTemplate: PromptFn<{
+      type: EventType;
+      jsonSchema: string;
+      context: string;
+      question: string;
+    }>,
     type: EventType,
     jsonSchema: any,
   ): ReaderTaskEither<C, APIError, PromptTemplate> =>
   (ctx) => {
     return fp.TE.tryCatch(async () => {
       const prompt = await PromptTemplate.fromTemplate(
-        promptTemplate({ type, jsonSchema }),
+        promptTemplate({
+          vars: {
+            type,
+            jsonSchema,
+            question: "{question}",
+            context: "{context}",
+          },
+        }),
       ).partial({
         evenType: type,
         jsonSchema: JSON.stringify(jsonSchema),
@@ -46,7 +58,12 @@ export const getCreateEventPromptPartial =
 export const createEventFromText = <C extends LoggerContext & LangchainContext>(
   text: Document[],
   type: EventType,
-  promptTemplate: PromptFn,
+  promptTemplate: PromptFn<{
+    jsonSchema: string;
+    type: EventType;
+    context: string;
+    question: string;
+  }>,
   jsonSchema: string,
   question: string,
 ): ReaderTaskEither<C, APIError | DBError, Event> => {

packages/@liexp/backend/src/providers/ai/langchain.provider.ts

@@ -7,49 +7,61 @@ import {
 } from "@langchain/core/runnables";
 import { ChatOpenAI, OpenAIEmbeddings } from "@langchain/openai";
 import { GetLogger } from "@liexp/core/lib/logger/index.js";
+import { type PromptFn } from "@liexp/shared/lib/io/openai/prompts/prompt.type.js";
 import type * as Reader from "fp-ts/lib/Reader.js";
 import { loadSummarizationChain } from "langchain/chains";
 import { RecursiveCharacterTextSplitter } from "langchain/text_splitter";
 import { formatDocumentsAsString } from "langchain/util/document";
 import { MemoryVectorStore } from "langchain/vectorstores/memory";
 
-export const EMBEDDINGS_PROMPT = `You are an assistant for question-answering tasks.
+export const EMBEDDINGS_PROMPT: PromptFn<{
+  text: string;
+  question: string;
+}> = ({
+  vars: { text, question },
+}) => `You are an assistant for question-answering tasks.
 Use the following pieces of retrieved context to answer the question.
 If you don't know the answer, just say that you don't know.
 Use 300 chars maximum and keep the answers concise.
 
 ---
-{context}
+${text}
 ---
 
-Question: {question}
+Question: ${question}
 
 Answer:
 `;
 
 // const DEFAULT_SUMMARIZATION_QUESTION = "Can you summarize this article for me?";
 
-export const DEFAULT_SUMMARIZE_PROMPT = `
+export const DEFAULT_SUMMARIZE_PROMPT: PromptFn<{ text: string }> = ({
+  vars: { text },
+}) => `
 You are an expert in summarizing texts. These texts can be either excerpt of web pages or articles.
 Your goal is to create a summary of the given text, focusing on the actions made by the characters mentioned in the text.
 Below you find the text you need to summarize.
 
 --------
-{text}
+${text}
 --------
 `;
 
 export interface LangchainProvider {
   chat: ChatOpenAI;
   embeddings: OpenAIEmbeddings;
-  queryDocument: (
+  queryDocument: <Args extends { text: string; question?: string }>(
     url: LangchainDocument[],
     question: string,
-    options?: { model?: AvailableModels; prompt?: string },
+    options?: { model?: AvailableModels; prompt?: PromptFn<Args> },
   ) => Promise<string>;
-  summarizeText: (
+  summarizeText: <Args extends { text: string }>(
     text: LangchainDocument[],
-    options?: { model?: AvailableModels; prompt?: string; question?: string },
+    options?: {
+      model?: AvailableModels;
+      prompt?: PromptFn<Args>;
+      question?: string;
+    },
   ) => Promise<string>;
 }
 
@@ -144,11 +156,13 @@ export const GetLangchainProvider = (
       // Retrieve and generate using the relevant snippets of the blog.
       const retriever = vectorStore.asRetriever();
 
-      const prompt = PromptTemplate.fromTemplate(EMBEDDINGS_PROMPT);
+      const prompt = PromptTemplate.fromTemplate(
+        EMBEDDINGS_PROMPT({ vars: { question: "{question}", text: "{text}" } }),
+      );
 
       const ragChain = RunnableSequence.from([
         {
-          context: retriever.pipe(formatDocumentsAsString),
+          text: retriever.pipe(formatDocumentsAsString),
           question: new RunnablePassthrough(),
         },
         prompt,
@@ -165,7 +179,14 @@ export const GetLangchainProvider = (
 
       return output;
     },
-    summarizeText: async (text, options) => {
+    summarizeText: async <Args extends { text: string }>(
+      text: LangchainDocument[],
+      options?: {
+        model?: AvailableModels;
+        prompt?: PromptFn<Args>;
+        question?: string;
+      },
+    ) => {
       const model = options?.model ?? opts.models?.chat ?? "gpt-4o";
       const prompt = options?.prompt ?? DEFAULT_SUMMARIZE_PROMPT;
 
@@ -185,7 +206,13 @@ export const GetLangchainProvider = (
       });
       const docsSummary = await textSplitter.splitDocuments(text);
 
-      const SUMMARY_PROMPT = PromptTemplate.fromTemplate(prompt);
+      const SUMMARY_PROMPT = PromptTemplate.fromTemplate(
+        prompt({
+          vars: {
+            text: docsSummary.flatMap((doc) => doc.pageContent).join("\n"),
+          } as Args,
+        }),
+      );
 
       const summarizeChain = loadSummarizationChain(chat, {
         type: "stuff",

packages/@liexp/shared/src/io/openai/prompts/actor.prompts.ts

@@ -1,23 +1,32 @@
 import { type PromptFn } from "./prompt.type.js";
 
-export const EMBED_ACTOR_PROMPT: PromptFn = () => `
+export const EMBED_ACTOR_PROMPT: PromptFn<{ text: string }> = ({
+  vars: { text },
+}) => `
 You are an expert in giving description about people.
-Your goal is to give a description of this person in a text format, including the requested fields, without inventing details.
+Your goal is to give a description of a given person in a text format, including the requested fields, without inventing details.
 The text should be minimum 100 words long, but not exceeding 300 words long.
 
 The requested fields are:
     - name: the name of the person
-    - birthOn: the birth date of the person in the format "YYYY-MM-DD"
-    - diedOn: the death date of the person in the format "YYYY-MM-DD"
+    - birthOn: the birth date of the person (in the format "YYYY-MM-DD")
+    - diedOn: the death date of the person (in the format "YYYY-MM-DD")
     - excerpt: a short description of the person (leave 2 empty lines at the end of this block)
     - body: a longer description of the person
 
-----
-{text}
-----
+Here's the name of the person you need to describe:
+
+---------------------------------------------------------------
+${text}
+---------------------------------------------------------------
 `;
 
-export const ACTOR_GENERAL_INFO_PROMPT: PromptFn = () => `
+export const ACTOR_GENERAL_INFO_PROMPT: PromptFn<{
+  text: string;
+  question: string;
+}> = ({ vars }) => {
+  const { text, question } = vars;
+  return `
 You are an expert in giving description about people.
 Your goal is to create a description including the requested fields about a given person, without inventing details.
 
@@ -30,11 +39,12 @@ The requested fields are:
 
 Below you find the person name and, optionally, a previous description of the person.
 
---------
-{text}
---------
+---------------------------------------------------------------
+${text}
+---------------------------------------------------------------
 
-Question: {question}
+Question: ${question}
 
-Answer:
+Give me your answer.
 `;
+};

packages/@liexp/shared/src/io/openai/prompts/event.prompts.ts

@@ -1,55 +1,63 @@
-import { EventType } from "../../http/Events/EventType.js";
+import { type EventType } from "../../http/Events/EventType.js";
 import { type PromptFn } from "./prompt.type.js";
 
-export const EMBED_EVENT_PROMPT = () => `
+export const EMBED_EVENT_PROMPT: PromptFn<{ text: string }> = ({ vars }) => `
 You are an expert in summarizing texts. These texts can be either excerpt of web pages or articles.
 Your goal is to create a summary of the given text, focusing on the actions made by the characters mentioned in the text.
 Try to extract the date of the event, the groups and actors involved.
+Respond with a professional summary of the event, without starting with "the text is about" or similar expression.
 
-Below you find the text you need to summarize.
-
---------
-{text}
---------
+---------------------------------------------------------------
+${vars.text}
+---------------------------------------------------------------
 `;
 
-const CREATE_EVENT_PROMPT = () => `
+const CREATE_EVENT_PROMPT: PromptFn<{
+  jsonSchema: string;
+  type: EventType;
+}> = ({ vars }) => `
 You are an expert in summarizing texts. These texts can be either excerpt of web pages or articles.
 Your goal is to create a summary of the given text, focusing on the actions made by the characters mentioned in given context.
 Try to match the dates with the given fields, using ISO 8601 format.
-The event can be one of the following type: ${EventType.types.map((t) => t.value).join(", ")}.
+The event type is the following: "${vars.type}".
 
 You return the summarized text in the "excerpt" key of the json object, and adapt others information to the following JSON OPENAPI schema:
 
----
-{jsonSchema}
----
+---------------------------------------------------------------
+${vars.jsonSchema}
+---------------------------------------------------------------
+
 `;
 
 /**
  * Prompt for creating an event from text.
  *
- * @param format_instructions Instructions on how to format the event.
- * @param text The text to extract the event from.
  */
-export const CREATE_EVENT_FROM_URL_PROMPT = () => `
-  ${CREATE_EVENT_PROMPT()}
+export const CREATE_EVENT_FROM_URL_PROMPT: PromptFn<{
+  jsonSchema: string;
+  context: string;
+  type: EventType;
+}> = ({ vars }) => `
+  ${CREATE_EVENT_PROMPT({ vars })}
 
   The context you need to extract the event from is:
 
-  --------
-  {context}
-  --------
+  ---------------------------------------------------------------
+  ${vars.context}
+  ---------------------------------------------------------------
 `;
 
 /**
- * Prompt for creating an event from a question.
+ * Prompt for creating an event from a text.
  *
- * @param eventType The type of event to create.
- * @param question The question to create the event from.
  */
-export const CREATE_EVENT_FROM_TEXT_PROMPT: PromptFn = () => `
-${CREATE_EVENT_FROM_URL_PROMPT()}
-
-Try to create a valid event, without inventing any detail from the following question: {question}
+export const CREATE_EVENT_FROM_TEXT_PROMPT: PromptFn<{
+  jsonSchema: string;
+  context: string;
+  type: EventType;
+  question: string;
+}> = ({ vars }) => `
+${CREATE_EVENT_FROM_URL_PROMPT({ vars })}
+
+Try to create a valid event, without inventing any detail from the following question: ${vars.question}
 `;

packages/@liexp/shared/src/io/openai/prompts/group.prompts.ts

@@ -1,6 +1,8 @@
 import { type PromptFn } from "./prompt.type.js";
 
-export const EMBED_GROUP_SUMMARIZE_PROMPT: PromptFn = () => `
+export const EMBED_GROUP_SUMMARIZE_PROMPT: PromptFn<{ text: string }> = ({
+  vars: { text },
+}) => `
 You are an expert in giving description about group.
 The group can be either a company or a website entity, a group of people, a family group.
 
@@ -16,8 +18,9 @@ The requested fields are:
 
 The text should be maximum 200 words long.
 
-----
-{text}
-----
+Here's the text for the group:
 
+---------------------------------------------------------------
+${text}
+---------------------------------------------------------------
 `;

packages/@liexp/shared/src/io/openai/prompts/link.prompts.ts

@@ -1,11 +1,13 @@
 import { type PromptFn } from "./prompt.type.js";
 
-export const EMBED_LINK_PROMPT: PromptFn = () => `
+export const EMBED_LINK_PROMPT: PromptFn<{ text: string }> = ({
+  vars: { text },
+}) => `
 You are an expert in summarizing texts. These texts can be either excerpt of web pages or articles.
-Your goal is to create a summary of the given text, focusing on the actions made by the characters mentioned in the text.
+Your goal is to create a summary of the given text, focusing on the actions made by the characters mentioned in the text and rephrasing using an investigative tone.
 Below you find the text you need to summarize.
 
---------
-{text}
---------
+---------------------------------------------------------------
+${text}
+---------------------------------------------------------------
 `;

packages/@liexp/shared/src/io/openai/prompts/prompt.type.ts

@@ -1 +1,3 @@
-export type PromptFn = (...args: any[]) => string;
+export type PromptFn<Vars = undefined> = Vars extends undefined
+  ? () => string
+  : (args: { vars: Vars }) => string;

services/ai-bot/src/flows/ai/createEventFromText.flow.ts

@@ -5,7 +5,7 @@ import { type CreateEventFromTextTypeData } from "@liexp/shared/lib/io/http/Queu
 import { toAIBotError } from "../../common/error/index.js";
 import { type ClientContext } from "../../context.js";
 import { loadDocs } from "./common/loadDocs.flow.js";
-import { getPromptFromResource } from "./prompts.js";
+import { getEventFromJsonPrompt } from "./prompts.js";
 import { type JobProcessRTE } from "#services/job-processor/job-processor.service.js";
 
 export const createEventFromTextFlow: JobProcessRTE<
@@ -30,7 +30,7 @@ export const createEventFromTextFlow: JobProcessRTE<
       if (job.prompt) {
         return fp.RTE.right(() => job.prompt!);
       }
-      return fp.RTE.right(getPromptFromResource(job.resource, job.type));
+      return fp.RTE.right(getEventFromJsonPrompt(job.type));
     }),
     fp.RTE.chain(({ docs, jsonSchema, prompt }) =>
       createEventFromText<ClientContext>(