remi/Perplexica
1
0
Fork 0
mirror of https://github.com/ItzCrazyKns/Perplexica.git synced 2025-10-31 19:38:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat(metaSearchAgent): eliminate runnables

Browse Source
This commit is contained in:
ItzCrazyKns
2025-03-24 17:27:54 +05:30
parent 45df9dc5bf
commit 0fcd598ff7
1 changed files with 215 additions and 254 deletions
Download Patch File Download Diff File Expand all files Collapse all files

469
src/lib/search/metaSearchAgent.ts
View File

@@ -6,11 +6,6 @@ import {
MessagesPlaceholder, MessagesPlaceholder,
PromptTemplate, PromptTemplate,
} from '@langchain/core/prompts'; } from '@langchain/core/prompts';
import {
RunnableLambda,
RunnableMap,
RunnableSequence,
} from '@langchain/core/runnables';
import { BaseMessage } from '@langchain/core/messages'; import { BaseMessage } from '@langchain/core/messages';
import { StringOutputParser } from '@langchain/core/output_parsers'; import { StringOutputParser } from '@langchain/core/output_parsers';
import LineListOutputParser from '../outputParsers/listLineOutputParser'; import LineListOutputParser from '../outputParsers/listLineOutputParser';
@@ -24,6 +19,7 @@ import computeSimilarity from '../utils/computeSimilarity';
import formatChatHistoryAsString from '../utils/formatHistory'; import formatChatHistoryAsString from '../utils/formatHistory';
import eventEmitter from 'events'; import eventEmitter from 'events';
import { StreamEvent } from '@langchain/core/tracers/log_stream'; import { StreamEvent } from '@langchain/core/tracers/log_stream';
import { EventEmitter } from 'node:stream';
export interface MetaSearchAgentType { export interface MetaSearchAgentType {
searchAndAnswer: ( searchAndAnswer: (
@@ -46,7 +42,7 @@ interface Config {
activeEngines: string[]; activeEngines: string[];
} }
type BasicChainInput = { type SearchInput = {
chat_history: BaseMessage[]; chat_history: BaseMessage[];
query: string; query: string;
}; };
@@ -59,235 +55,240 @@ class MetaSearchAgent implements MetaSearchAgentType {
this.config = config; this.config = config;
} }
private async createSearchRetrieverChain(llm: BaseChatModel) { private async searchSources(
llm: BaseChatModel,
input: SearchInput,
emitter: EventEmitter,
) {
(llm as unknown as ChatOpenAI).temperature = 0; (llm as unknown as ChatOpenAI).temperature = 0;
return RunnableSequence.from([ const chatPrompt = PromptTemplate.fromTemplate(
PromptTemplate.fromTemplate(this.config.queryGeneratorPrompt), this.config.queryGeneratorPrompt,
llm, );
this.strParser,
RunnableLambda.from(async (input: string) => {
const linksOutputParser = new LineListOutputParser({
key: 'links',
});
const questionOutputParser = new LineOutputParser({ const processedChatPrompt = await chatPrompt.invoke({
key: 'question', chat_history: formatChatHistoryAsString(input.chat_history),
}); query: input.query,
});
const links = await linksOutputParser.parse(input); const llmRes = await llm.invoke(processedChatPrompt);
let question = this.config.summarizer const messageStr = await this.strParser.invoke(llmRes);
? await questionOutputParser.parse(input)
: input;
if (question === 'not_needed') { const linksOutputParser = new LineListOutputParser({
return { query: '', docs: [] }; key: 'links',
});
const questionOutputParser = new LineOutputParser({
key: 'question',
});
const links = await linksOutputParser.parse(messageStr);
let question = this.config.summarizer
? await questionOutputParser.parse(messageStr)
: messageStr;
if (question === 'not_needed') {
return { query: '', docs: [] };
}
if (links.length > 0) {
if (question.length === 0) {
question = 'summarize';
}
let docs: Document[] = [];
const linkDocs = await getDocumentsFromLinks({ links });
const docGroups: Document[] = [];
linkDocs.map((doc) => {
const URLDocExists = docGroups.find(
(d) =>
d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
);
if (!URLDocExists) {
docGroups.push({
...doc,
metadata: {
...doc.metadata,
totalDocs: 1,
},
});
} }
if (links.length > 0) { const docIndex = docGroups.findIndex(
if (question.length === 0) { (d) =>
question = 'summarize'; d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
} );
let docs: Document[] = []; if (docIndex !== -1) {
docGroups[docIndex].pageContent =
const linkDocs = await getDocumentsFromLinks({ links }); docGroups[docIndex].pageContent + `\n\n` + doc.pageContent;
docGroups[docIndex].metadata.totalDocs += 1;
const docGroups: Document[] = [];
linkDocs.map((doc) => {
const URLDocExists = docGroups.find(
(d) =>
d.metadata.url === doc.metadata.url &&
d.metadata.totalDocs < 10,
);
if (!URLDocExists) {
docGroups.push({
...doc,
metadata: {
...doc.metadata,
totalDocs: 1,
},
});
}
const docIndex = docGroups.findIndex(
(d) =>
d.metadata.url === doc.metadata.url &&
d.metadata.totalDocs < 10,
);
if (docIndex !== -1) {
docGroups[docIndex].pageContent =
docGroups[docIndex].pageContent + `\n\n` + doc.pageContent;
docGroups[docIndex].metadata.totalDocs += 1;
}
});
await Promise.all(
docGroups.map(async (doc) => {
const res = await llm.invoke(`
You are a web search summarizer, tasked with summarizing a piece of text retrieved from a web search. Your job is to summarize the
text into a detailed, 2-4 paragraph explanation that captures the main ideas and provides a comprehensive answer to the query.
If the query is \"summarize\", you should provide a detailed summary of the text. If the query is a specific question, you should answer it in the summary.
- **Journalistic tone**: The summary should sound professional and journalistic, not too casual or vague.
- **Thorough and detailed**: Ensure that every key point from the text is captured and that the summary directly answers the query.
- **Not too lengthy, but detailed**: The summary should be informative but not excessively long. Focus on providing detailed information in a concise format.
The text will be shared inside the \`text\` XML tag, and the query inside the \`query\` XML tag.
<example>
1. \`<text>
Docker is a set of platform-as-a-service products that use OS-level virtualization to deliver software in packages called containers.
It was first released in 2013 and is developed by Docker, Inc. Docker is designed to make it easier to create, deploy, and run applications
by using containers.
</text>
<query>
What is Docker and how does it work?
</query>
Response:
Docker is a revolutionary platform-as-a-service product developed by Docker, Inc., that uses container technology to make application
deployment more efficient. It allows developers to package their software with all necessary dependencies, making it easier to run in
any environment. Released in 2013, Docker has transformed the way applications are built, deployed, and managed.
\`
2. \`<text>
The theory of relativity, or simply relativity, encompasses two interrelated theories of Albert Einstein: special relativity and general
relativity. However, the word "relativity" is sometimes used in reference to Galilean invariance. The term "theory of relativity" was based
on the expression "relative theory" used by Max Planck in 1906. The theory of relativity usually encompasses two interrelated theories by
Albert Einstein: special relativity and general relativity. Special relativity applies to all physical phenomena in the absence of gravity.
General relativity explains the law of gravitation and its relation to other forces of nature. It applies to the cosmological and astrophysical
realm, including astronomy.
</text>
<query>
summarize
</query>
Response:
The theory of relativity, developed by Albert Einstein, encompasses two main theories: special relativity and general relativity. Special
relativity applies to all physical phenomena in the absence of gravity, while general relativity explains the law of gravitation and its
relation to other forces of nature. The theory of relativity is based on the concept of "relative theory," as introduced by Max Planck in
1906. It is a fundamental theory in physics that has revolutionized our understanding of the universe.
\`
</example>
Everything below is the actual data you will be working with. Good luck!
<query>
${question}
</query>
<text>
${doc.pageContent}
</text>
Make sure to answer the query in the summary.
`);
const document = new Document({
pageContent: res.content as string,
metadata: {
title: doc.metadata.title,
url: doc.metadata.url,
},
});
docs.push(document);
}),
);
return { query: question, docs: docs };
} else {
question = question.replace(/<think>.*?<\/think>/g, '');
const res = await searchSearxng(question, {
language: 'en',
engines: this.config.activeEngines,
});
const documents = res.results.map(
(result) =>
new Document({
pageContent:
result.content ||
(this.config.activeEngines.includes('youtube')
? result.title
: '') /* Todo: Implement transcript grabbing using Youtubei (source: https://www.npmjs.com/package/youtubei) */,
metadata: {
title: result.title,
url: result.url,
...(result.img_src && { img_src: result.img_src }),
},
}),
);
return { query: question, docs: documents };
} }
}), });
]);
await Promise.all(
docGroups.map(async (doc) => {
const res = await llm.invoke(`
You are a web search summarizer, tasked with summarizing a piece of text retrieved from a web search. Your job is to summarize the
text into a detailed, 2-4 paragraph explanation that captures the main ideas and provides a comprehensive answer to the query.
If the query is \"summarize\", you should provide a detailed summary of the text. If the query is a specific question, you should answer it in the summary.
- **Journalistic tone**: The summary should sound professional and journalistic, not too casual or vague.
- **Thorough and detailed**: Ensure that every key point from the text is captured and that the summary directly answers the query.
- **Not too lengthy, but detailed**: The summary should be informative but not excessively long. Focus on providing detailed information in a concise format.
The text will be shared inside the \`text\` XML tag, and the query inside the \`query\` XML tag.
<example>
1. \`<text>
Docker is a set of platform-as-a-service products that use OS-level virtualization to deliver software in packages called containers.
It was first released in 2013 and is developed by Docker, Inc. Docker is designed to make it easier to create, deploy, and run applications
by using containers.
</text>
<query>
What is Docker and how does it work?
</query>
Response:
Docker is a revolutionary platform-as-a-service product developed by Docker, Inc., that uses container technology to make application
deployment more efficient. It allows developers to package their software with all necessary dependencies, making it easier to run in
any environment. Released in 2013, Docker has transformed the way applications are built, deployed, and managed.
\`
2. \`<text>
The theory of relativity, or simply relativity, encompasses two interrelated theories of Albert Einstein: special relativity and general
relativity. However, the word "relativity" is sometimes used in reference to Galilean invariance. The term "theory of relativity" was based
on the expression "relative theory" used by Max Planck in 1906. The theory of relativity usually encompasses two interrelated theories by
Albert Einstein: special relativity and general relativity. Special relativity applies to all physical phenomena in the absence of gravity.
General relativity explains the law of gravitation and its relation to other forces of nature. It applies to the cosmological and astrophysical
realm, including astronomy.
</text>
<query>
summarize
</query>
Response:
The theory of relativity, developed by Albert Einstein, encompasses two main theories: special relativity and general relativity. Special
relativity applies to all physical phenomena in the absence of gravity, while general relativity explains the law of gravitation and its
relation to other forces of nature. The theory of relativity is based on the concept of "relative theory," as introduced by Max Planck in
1906. It is a fundamental theory in physics that has revolutionized our understanding of the universe.
\`
</example>
Everything below is the actual data you will be working with. Good luck!
<query>
${question}
</query>
<text>
${doc.pageContent}
</text>
Make sure to answer the query in the summary.
`);
const document = new Document({
pageContent: res.content as string,
metadata: {
title: doc.metadata.title,
url: doc.metadata.url,
},
});
docs.push(document);
}),
);
return { query: question, docs: docs };
} else {
question = question.replace(/<think>.*?<\/think>/g, '');
const res = await searchSearxng(question, {
language: 'en',
engines: this.config.activeEngines,
});
const documents = res.results.map(
(result) =>
new Document({
pageContent:
result.content ||
(this.config.activeEngines.includes('youtube')
? result.title
: '') /* Todo: Implement transcript grabbing using Youtubei (source: https://www.npmjs.com/package/youtubei) */,
metadata: {
title: result.title,
url: result.url,
...(result.img_src && { img_src: result.img_src }),
},
}),
);
return { query: question, docs: documents };
}
} }
private async createAnsweringChain( private async streamAnswer(
llm: BaseChatModel, llm: BaseChatModel,
fileIds: string[], fileIds: string[],
embeddings: Embeddings, embeddings: Embeddings,
optimizationMode: 'speed' | 'balanced' | 'quality', optimizationMode: 'speed' | 'balanced' | 'quality',
input: SearchInput,
emitter: EventEmitter,
) { ) {
return RunnableSequence.from([ const chatPrompt = ChatPromptTemplate.fromMessages([
RunnableMap.from({ ['system', this.config.responsePrompt],
query: (input: BasicChainInput) => input.query, new MessagesPlaceholder('chat_history'),
chat_history: (input: BasicChainInput) => input.chat_history, ['user', '{query}'],
date: () => new Date().toISOString(), ]);
context: RunnableLambda.from(async (input: BasicChainInput) => {
const processedHistory = formatChatHistoryAsString(
input.chat_history,
);
let docs: Document[] | null = null; let docs: Document[] | null = null;
let query = input.query; let query = input.query;
if (this.config.searchWeb) { if (this.config.searchWeb) {
const searchRetrieverChain = const searchResults = await this.searchSources(llm, input, emitter);
await this.createSearchRetrieverChain(llm);
const searchRetrieverResult = await searchRetrieverChain.invoke({ query = searchResults.query;
chat_history: processedHistory, docs = searchResults.docs;
query, }
});
query = searchRetrieverResult.query; const sortedDocs = await this.rerankDocs(
docs = searchRetrieverResult.docs; query,
} docs ?? [],
fileIds,
embeddings,
optimizationMode,
);
const sortedDocs = await this.rerankDocs( emitter.emit('data', JSON.stringify({ type: 'sources', data: sortedDocs }));
query,
docs ?? [],
fileIds,
embeddings,
optimizationMode,
);
return sortedDocs; const context = this.processDocs(sortedDocs);
})
.withConfig({ const formattedChatPrompt = await chatPrompt.invoke({
runName: 'FinalSourceRetriever', query: input.query,
}) chat_history: input.chat_history,
.pipe(this.processDocs), date: new Date().toISOString(),
}), context: context,
ChatPromptTemplate.fromMessages([
['system', this.config.responsePrompt],
new MessagesPlaceholder('chat_history'),
['user', '{query}'],
]),
llm,
this.strParser,
]).withConfig({
runName: 'FinalResponseGenerator',
}); });
const llmRes = await llm.stream(formattedChatPrompt);
for await (const data of llmRes) {
const messageStr = await this.strParser.invoke(data);
emitter.emit(
'data',
JSON.stringify({ type: 'response', data: messageStr }),
);
}
emitter.emit('end');
} }
private async rerankDocs( private async rerankDocs(
@@ -428,39 +429,6 @@ class MetaSearchAgent implements MetaSearchAgentType {
.join('\n'); .join('\n');
} }
private async handleStream(
stream: AsyncGenerator<StreamEvent, any, any>,
emitter: eventEmitter,
) {
for await (const event of stream) {
if (
event.event === 'on_chain_end' &&
event.name === 'FinalSourceRetriever'
) {
``;
emitter.emit(
'data',
JSON.stringify({ type: 'sources', data: event.data.output }),
);
}
if (
event.event === 'on_chain_stream' &&
event.name === 'FinalResponseGenerator'
) {
emitter.emit(
'data',
JSON.stringify({ type: 'response', data: event.data.chunk }),
);
}
if (
event.event === 'on_chain_end' &&
event.name === 'FinalResponseGenerator'
) {
emitter.emit('end');
}
}
}
async searchAndAnswer( async searchAndAnswer(
message: string, message: string,
history: BaseMessage[], history: BaseMessage[],
@@ -471,25 +439,18 @@ class MetaSearchAgent implements MetaSearchAgentType {
) { ) {
const emitter = new eventEmitter(); const emitter = new eventEmitter();
const answeringChain = await this.createAnsweringChain( this.streamAnswer(
llm, llm,
fileIds, fileIds,
embeddings, embeddings,
optimizationMode, optimizationMode,
);
const stream = answeringChain.streamEvents(
{ {
chat_history: history, chat_history: history,
query: message, query: message,
}, },
{ emitter,
version: 'v1',
},
); );
this.handleStream(stream, emitter);
return emitter; return emitter;
} }
} }