2023-09-09 Asya.AI Chatbot ShowCase

 

Demo usecases asya.ai chatbots

  1. Latviešu valodā 1230 fakti http://tele2.asya.ai

  2. Latviešu valodā 545 fakti http://tet.asya.ai

  1. Ierobežojumi (demo)

    1. Viena čata ietvaros lūdzu runāt par vienu tēmu, apakšā ir poga ar kuru sākt jaunu tēmu. Pašlaik sistēma cenšas atpazīt, ja tēma mainījusies, bet tā nav perfekta.

    1. Sistēma šobrīd ir lēna, jo izmanto openai modeļus, bet to var paātrināt, ja pārejam uz mūsu modeļiem, kad esam noslēguši darījumu.

    1. Šobrīd nav ieteicams vairāk par 5 papildjautājumiem vienas tēmas ietvaros

    1. Jautājumi līdz 100 vārdiem šobrīd

 

image-20230909130458511

 

image-20230909130548445

 

image-20230909130609825

image-20230909130628670

image-20230909130643412

image-20230909130700658

 

 

image-20230909130719354

 

 

Comparison of ChatBots

Based on text similarity between ground truth chat and generated chat messages

Comparison of ChatBot scores (max 10)7.92 7.36 7.63 OpenAI APIPrivateGPTDelloite Comparison of ChatBot scores (max 10)
Comparison of ChatBot accuracy87.5% 77.5% 90.0% OpenAI APIPrivateGPTDelloite Comparison of ChatBot accuracy
ChatBot embeddings for extracting meaning (smaller distance better)0.35 0.21 0.16 0.12 0.12 0.09 alfaneoOpenAI APIjinja_embmulti_miniLM_L12all_mpnet_basesmlr ChatBot embeddings for extracting meaning (smaller distance better)
ChatBot embeddings for extracting structure (larger distance better)0.39 0.59 0.59 0.60 0.61 0.59 alfaneoOpenAI APIjinja_embmulti_miniLM_L12all_mpnet_basesmlr ChatBot embeddings for extracting structure (larger distance better)

 

 

Chatbot system architecture

Compute
UX
Datasource formatting API
Text embedding API
LLM API
Chatbot API
Datasorce provider API
Vector database
Exasol
Data source
PDF, DOCX
Prior conversions
Client databases
REST API
Web page
Mobile app
Email
Whatsapp, IM
Admin app

Explanations:

  1. Data Source

    1. Existing client data that contains information needed to handle

  2. Compute

    1. Datasource formatting API - Takes data sources like PDFs, prior conversations and convert them into vector database (need to be adapted to customer needs)

    2. Datasorce provider API - Provides Chatbot API with necessary information based on vector queries. This abstraction to ensure system can work with any Vector Databse - Exasol, PostgreSQL, Oracle etc. based on client needs

    3. Text embedding API - Allows to convert natural language text into vector embeddings

    4. LLM API - Large Language Model trained to be InstructGPT based model

    5. Chatbot API - Main chatbot logic (need to be adapted to customer needs)

  3. Vector Database

    1. Any database to store vector and text chunks and other information like images, videos, sensor data etc.

  4. UX - user experience for chatbot (can be actual app or just API)

    1. Admin App - Special tool to see chatbot statistics and conversation flows, valuable conversational intelligence. Allows to modify chatbot behaviour in real time.