8b9b4085ea
The initial query appears twice in the prompt, this is ignored by OpenAI models, however it breaks with Gemini models are they fail with an error stating that AI and User prompts need to alternate. Tested all search modes with both OpenAI GTP3 turbo and Vertex Gemini 1.0 and this changes appears to now function correctly with both.
89 lines
2.6 KiB
TypeScript
89 lines
2.6 KiB
TypeScript
import { BaseMessage } from '@langchain/core/messages';
|
|
import {
|
|
ChatPromptTemplate,
|
|
MessagesPlaceholder,
|
|
} from '@langchain/core/prompts';
|
|
import { RunnableSequence } from '@langchain/core/runnables';
|
|
import { StringOutputParser } from '@langchain/core/output_parsers';
|
|
import type { StreamEvent } from '@langchain/core/tracers/log_stream';
|
|
import eventEmitter from 'events';
|
|
import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
|
|
import type { Embeddings } from '@langchain/core/embeddings';
|
|
import logger from '../utils/logger';
|
|
|
|
const writingAssistantPrompt = `
|
|
You are Perplexica, an AI model who is expert at searching the web and answering user's queries. You are currently set on focus mode 'Writing Assistant', this means you will be helping the user write a response to a given query.
|
|
Since you are a writing assistant, you would not perform web searches. If you think you lack information to answer the query, you can ask the user for more information or suggest them to switch to a different focus mode.
|
|
`;
|
|
|
|
const strParser = new StringOutputParser();
|
|
|
|
const handleStream = async (
|
|
stream: AsyncGenerator<StreamEvent, any, unknown>,
|
|
emitter: eventEmitter,
|
|
) => {
|
|
for await (const event of stream) {
|
|
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');
|
|
}
|
|
}
|
|
};
|
|
|
|
const createWritingAssistantChain = (llm: BaseChatModel) => {
|
|
return RunnableSequence.from([
|
|
ChatPromptTemplate.fromMessages([
|
|
['system', writingAssistantPrompt],
|
|
new MessagesPlaceholder('chat_history'),
|
|
]),
|
|
llm,
|
|
strParser,
|
|
]).withConfig({
|
|
runName: 'FinalResponseGenerator',
|
|
});
|
|
};
|
|
|
|
const handleWritingAssistant = (
|
|
query: string,
|
|
history: BaseMessage[],
|
|
llm: BaseChatModel,
|
|
embeddings: Embeddings,
|
|
) => {
|
|
const emitter = new eventEmitter();
|
|
|
|
try {
|
|
const writingAssistantChain = createWritingAssistantChain(llm);
|
|
const stream = writingAssistantChain.streamEvents(
|
|
{
|
|
chat_history: history,
|
|
query: query,
|
|
},
|
|
{
|
|
version: 'v1',
|
|
},
|
|
);
|
|
|
|
handleStream(stream, emitter);
|
|
} catch (err) {
|
|
emitter.emit(
|
|
'error',
|
|
JSON.stringify({ data: 'An error has occurred please try again later' }),
|
|
);
|
|
logger.error(`Error in writing assistant: ${err}`);
|
|
}
|
|
|
|
return emitter;
|
|
};
|
|
|
|
export default handleWritingAssistant;
|