Overview

This pipeline is designed to handle conversations with bank customers about missed auto loan payments. It's a smart customer service system that knows what to do next based on the conversation flow. It automatically chooses the right action based on context while keeping the conversation flowing naturally. It also always maintains a conversation record to stay aware of the context.

Best For

Automating customer service interactions
Analyzing sentiment and financial details to guide payment negotiations
Generating human-like prompts for client communication
Verifying client information with database integration
Handling payment instructions and escalation scenarios

Recommended Dataset

Any Snowflake database with an associated schema of tables and columns extracted from DESCRIBE TABLE. For details, see Snowflake documentation.
Any text data meant to assist with the chatbot's responses.

Prerequisites

This pipeline uses the gpt-4o model so you'll need an OpenAI API key to use it. (You can also change the Generator's model).
Table descriptions for retrieval:
1. For each table, create a TXT file with its description. The description should be short, one or two sentences is enough. This is meant to help the Retriever find the relevant table.
2. Add the following metadata to each TXT file: table: table_name" and "table_columns: column_name - data_type.
  For example, for a table called COUNTRY, you'd create a description file called COUNTRY.TXT with the following metadata: {"table": "COUNTRY", "table_columns": "ID - INTEGER\nNAME - VARCHAR\nREGION - VARCHAR"}.
3. Upload table descriptions to deepset Cloud to the same workspace where the pipeline is. For details, see Upload Files.

Potential Applications

Automated customer support system.

How It Works

The system analyzes the conversation and decides what action to take. It can choose from seven paths: start call, user verification, sentiment analysis, information extraction, negotiation logic, payment handling, or escalation and handoff.
Components involved: root_prompt_builder, root_llm, conditional_router
It then proceeds with handling the selected action:
1. Start call: Creates the initial greeting and asks for verification information.
  Components involved:
  inititate_call_prompt_builder, initiate_call_llm, initiate_call_answer_builder
  1. User verification: Verifies the information the customer provided against a Snowflake database. Then, it either confirms their identity or asks for correct information.
    Components involved:
    verification_prompt_builder, verification_llm, verification_success_adapter, client_database, verification_router, verification_success_prompt_builder, verification_success_llm, verification_success_answer_builder, verification_fail_prompt_builder, verification_fail_adapter, verification_fail_answer_builder
2. Sentiment analysis: Reads the customer's messages to understand their mood and adjusts the conversation tone accordingly.
  Components involved:
  sentiment_prompt_builder, sentiment_llm, sentiment_answer_builder
3. Information extraction: Focuses on details about the customer's financial situation, such as when they can pay or how much they can afford.
  Components involved:
  info_extract_prompt_builder, info_extract_llm, info_extract_answer_builder
4. Negotiation logic: Handles discussions about payment options and tries to find solutions that work for both the bank and the customer.
  Components involved:
  negotiation_prompt_builder, negotiation_llm, negotiation_answer_builder
5. Payment handling: Processes the actual payment arrangements and provides payment links.
  Components involved:
  payment_prompt_builder, payment_llm, payment_answer_builder
6. Escalation: Handles situations where a human needs to take over and ensures a smooth handoff to human support.
  Components involved:
  escalation_prompt_builder, escalation_llm, escalation_answer_builder

Pipeline YAML

  components:
    root_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are part of a chatbot. You receive a question (Current Question) and a chat history.
                  
          Analyze the following chat history between a client and deepset bank and decide the most appropriate action to take next. 
          The response may include indications of the client's sentiment, willingness to pay, or specific details about their financial situation. 
          Based on the chat history, select one of the following actions as the next step in the process:
          Start: If the client has not been authenticated. This should always be the first step.
          User Verification: If the client provides their first name, last name and account number for verification.
          Sentiment Analysis: If sentiment cues (e.g., frustration, relief) need to be evaluated to guide the negotiation tone.
          Information Extraction: If specific financial details (e.g., the ability to pay, amount, timing, or payment preferences) need to be identified.
          Negotiation Logic: If the client shows willingness to negotiate or requests information about possible payment plans.
          Payment Handling: If the client agrees to make a payment immediately.
          Escalation & Handoff: If the client expresses frustration, asks for human assistance, or the conversation has not progressed.
          Your answer must be one of the following actions:
            - Start Call
            - User Verification
            - Sentiment Analysis
            - Information Extraction
            - Negotiation Logic
            - Payment Handling
            - Escalation & Handoff
          Context: {{ question }}
          Next action:
    root_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        api_key:
          type: env_var
          env_vars:
            - OPENAI_API_KEY
          strict: false
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 650
          temperature: 0.0
          seed: 0
    conditional_router:  # Conditional router to determine the next action based on the client's response
      type: haystack.components.routers.conditional_router.ConditionalRouter
      init_parameters:
        routes:
          - condition: '{{replies[0].lower() == ''start call''}}'
            output: '{{ question }}'
            output_name: call
            output_type: str
          - condition: '{{replies[0].lower() == ''user verification''}}'
            output: '{{ question }}'
            output_name: verification
            output_type: str
          - condition: '{{replies[0].lower() == ''sentiment analysis''}}'
            output: '{{ question }}'
            output_name: sentiment
            output_type: str
          - condition: '{{replies[0].lower() == ''negotiation logic''}}'
            output: '{{ question }}'
            output_name: negotiation
            output_type: str
          - condition: '{{replies[0].lower() == ''escalation & handoff''}}'
            output: '{{ question }}'
            output_name: escalation
            output_type: str
          - condition: '{{replies[0].lower() == ''information extraction''}}'
            output: '{{ question }}'
            output_name: extract
            output_type: str
          - condition: '{{replies[0].lower() == ''payment handling''}}'
            output: '{{ question }}'
            output_name: payment
            output_type: str
        unsafe: false
    # Components for initiating the call
    initiate_call_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a chatbot calling a client to verify their account numbers before discussing missed auto loan payments. 
          Your name is Deepset, and you work for Deepset Bank as a customer service agent.
         
          Instructions:
          - Your role is to generate a prompt asking the client to verify their first and last name, along with their account number.
          - This is a one-way discussion; the only questions should be to verify the account number and the client’s first and last name.
          - Make sure you sound as human as possible when generating the prompt.
          {{ question }}
    initiate_call_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 1000
          temperature: 0.0
          seed: 0
    initiate_call_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    # Authentification components
    verification_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are an SQL expert. Your task is to extract the client's first name, last name, and account number from the following text {{ text }}. After identifying the name and account number, add them to the WHERE clause in the SQL template below. If you cannot identify the client's name in this text, respond with an empty string "". Your only goal is to provide the SQL template with the appropriate name or an empty string without any explanation.
          SQL template:
          SELECT * FROM CAR_LOAN_DB.WRITEOFF.CUSTOMERS
          WHERE CLIENT_NAME = <add name here>
          AND ACCOUNT_NUMBER = <add account number here>
          LIMIT 1;
          You must provide only the SQL query in plain text, as non-SQL characters will break the database. 
          Your answer must not include ```sql or ```. If you cannot identify the client in this text, respond with an empty string "".
    
    verification_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 1000
          temperature: 0.0
          seed: 0
    
    client_database:  # This is meant to access client information stored in Snowflake, from which the LLM will verify the information that the customer provided.
      type: deepset_cloud_custom_nodes.retrievers.snowflake_retriever.DeepsetSnowflakeRetriever
      init_parameters:
        user: "<snowflake-user-identifier>"
        account: "<snowflake-account-identifier>"
        warehouse: "<snowflake-warehouse-name>"
    
    verification_success_adapter:
      type: haystack.components.converters.output_adapter.OutputAdapter
      init_parameters:
        template: '{{ replies[0] }}'
        output_type: str
    
    verification_router:
      type: haystack.components.routers.conditional_router.ConditionalRouter
      init_parameters:
        routes:
          - condition: '{{ dataframe.shape[0] == 0 }}'
            output: '{{ table }}'
            output_name: fail
            output_type: str
          - condition: '{{ dataframe.shape[0] > 0 }}'
            output: '{{ table }}'
            output_name: success
            output_type: str
        unsafe: false
    
    verification_fail_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          I'm sorry, but I couldn't find a match for your first name, last name, and account number in our system. 
          Could you please double-check the information and provide it again? This will help us ensure we're connecting to the correct account. 
          Thank you for your patience!
          {{ table }}
    
    verification_fail_adapter:
      type: haystack.components.converters.output_adapter.OutputAdapter
      init_parameters:
        template: '{{ [prompt] }}'
        output_type: typing.List[str]
    
    verification_success_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a customer service agent at Deepset Bank responsible for negotiating auto loan payments with clients who have missed their payments. 
          Your task is to verify client details against the bank's records and start the discussion around paying the missed car loan payments. 
          Use the chat history below for guidance and context.
          Instructions:
          - Verify the user's identity by confirming their first name, last name, and account number.
          - Extract critical financial information from the table and ask the user if they would like to make the missed payment.
          - Provide them with an update on their loan, including details about the car, the loan amount, and the remaining balance to pay off the loan.
          - Remember, the generated prompt will be used for a text-to-speech system, so you should sound like a human.
          It's important to consider that the client you are discussing with has missed their payment.
          Bank's records: {{ table }}
          Chat history: {{ question }}
          Prompt:
    
    verification_success_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 2000
          temperature: 0.0
          seed: 0
    
    verification_fail_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    
    verification_success_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    
    verification_answer_builder_joiner:
      type: haystack.components.joiners.answer_joiner.AnswerJoiner
      init_parameters: {}
    
    # Components for detecting clients' sentiment
    sentiment_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a customer service agent at Deepset Bank responsible for negotiating auto loan payments with clients who have missed their payments. 
          You have been provided with the chat history and the client's current question below to help you achieve your goal. 
          
          Instructions:
            - Analyze the client's response in the chat history for emotional cues. 
            - Assess whether the sentiment is positive, neutral, or negative, and determine if the client is receptive, resistant, frustrated, or cooperative regarding commitment to payment or a payment arrangement. 
          
          Remember, the generated prompt will be used for a text-to-speech system, so you should sound like a human.
          It's important to consider that the client you are discussing with has missed their payment. 
          It is also important to remember that you are already on the call with the client. 
          The chat history below refers to the previous conversation. That said, do not act as if this is the first time you are talking to the client.
          
          {{ question }}
          
          Prompt:
    
    sentiment_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 2000
          temperature: 0.0
          seed: 0
    
    sentiment_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    
    # components for conducting the payment negotiation
    negotiation_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a customer service agent at Deepset Bank responsible for negotiating auto loan payments with clients who have missed their payments. 
          You have been provided with the chat history and the client's current question below to help you achieve your goal. 
          
          Instructions:
          - Generate a response based on the client’s sentiment and willingness to discuss payment options, aiming to reach a mutually agreeable payment solution. 
          - Adjust your tone according to the client’s responses. The response should maintain a respectful tone and be solution-oriented, encouraging the client toward a payment commitment that meets their current capacity.
          It's important to consider that the client you are discussing with has missed their payment. 
          Remember that you are already on the call with the client. The chat history below refers to the previous conversation, so do not act as if this is the first time you are talking to the client.
          
          {{ question }}
          
          Prompt:
    
    negotiation_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 2000
          temperature: 0.0
          seed: 0
    
    negotiation_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    
    # Components for escalations and handoffs
    escalation_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a customer service agent at Deepset Bank responsible for negotiating auto loan payments with clients who have missed their payments.
          You have been provided with the chat history and the client current question below to help you achieve your goal.
          
          Instructions:
          - The client’s response suggests they may need additional assistance from a human representative. 
          - Based on the tone of their message (e.g., frustration, confusion, or request for direct help), generate a response that politely acknowledges their concerns and reassures them that they will be connected with a team member who can help. 
          - Ensure the message conveys empathy and appreciation for their patience, maintaining a supportive and professional tone during the handoff."
          Example response:
          "Thank you for your patience, [Client’s Name].
          I understand that this might be a bit complex, so I’m connecting you with a team member who can assist further. 
          We appreciate you working with us, and someone will be with you shortly to make sure your concerns are addressed."
          It's important to consider that the client you are in discussion with has missed their payment.
          It is also important to remember that you are already on the call with the client. The chat history below refers to the previous conversation. That said, do not act as if this is the first time you are talking to the client.
          
          {{ question }}
          
          Prompt:
    
    escalation_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 2000
          temperature: 0.0
          seed: 0
    
    escalation_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    
    # Components for extracting relevant financial details 
    info_extract_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a customer service agent at Deepset Bank, responsible for negotiating auto loan payments with clients who have missed their payments. 
          You have been provided with the chat history and the client's current question below to help you achieve your goal. 
          
          Instructions: 
          - Analyze the client’s response and extract relevant financial details that can guide the negotiation process.
          - Look for information such as the client's ability or willingness to pay, along with specific payments. 
          
          It is important to consider that the client you are discussing with has missed their payment. It is also essential to remember that you are already on the call with the client. 
          The chat history below refers to the previous conversation. That said, do not act as if this is the first time you are talking to the client. 
          
          Example of extracted information: "The client indicated they can make a partial payment of $100 next week but cannot commit to the full balance at this time."
          
          {{ question }}
          
          Prompt:
    
    info_extract_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 2000
          temperature: 0.0
          seed: 0
    
    info_extract_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    
    # Component for handling the payment
    payment_prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a customer service agent at deepset Bank responsible for negotiating auto loan payments with clients who have missed their payments.
          You have been provided with the chat history and the client current question below to help you achieve your goal.
          
          Instructions:
          - Based on the client’s agreement to make a payment or partial payment, generate a message that confirms their willingness and provides instructions for completing the payment. 
          - If they have chosen a specific payment plan, outline the details clearly. Include a secure payment link, and thank the client for their cooperation. 
          - Ensure the tone is positive and appreciative, encouraging a smooth completion of the payment process."
          Payment Link: "https://www.deepset.ai"
          Example response:
          "Thank you for your cooperation, [Client’s Name]! You can complete your payment by following this secure link: [Payment Link]. If you have any questions, please let us know—we’re here to help!”
          It is also important to remember that you are already on the call with the client. 
          The chat history below refers to the previous conversation. That said, do not act as if this is the first time you are talking to the client.
          
          {{ question }}
          
          Prompt:
    
    payment_llm:
      type: haystack.components.generators.openai.OpenAIGenerator
      init_parameters:
        model: "gpt-4o"
        generation_kwargs:
          max_tokens: 2000
          temperature: 0.0
          seed: 0
    
    payment_answer_builder:
      type: deepset_cloud_custom_nodes.augmenters.deepset_answer_builder.DeepsetAnswerBuilder
      init_parameters:
        reference_pattern: acm
    
    answer_joiner:
      type: haystack.components.joiners.answer_joiner.AnswerJoiner
      init_parameters: {}
  connections:
    - sender: root_prompt_builder.prompt
      receiver: root_llm.prompt
    - sender: root_llm.replies
      receiver: conditional_router.replies
    - sender: conditional_router.call
      receiver: initiate_call_prompt_builder.question
    - sender: initiate_call_prompt_builder.prompt
      receiver: initiate_call_llm.prompt
    - sender: initiate_call_llm.replies
      receiver: initiate_call_answer_builder.replies
    - sender: conditional_router.call
      receiver: initiate_call_answer_builder.query
    - sender: conditional_router.verification
      receiver: verification_prompt_builder.text
    - sender: verification_prompt_builder.prompt
      receiver: verification_llm.prompt
    - sender: verification_llm.replies
      receiver: verification_success_adapter.replies
    - sender: verification_success_adapter.output
      receiver: client_database.query
    - sender: client_database.table
      receiver: verification_router.table
    - sender: verification_router.fail
      receiver: verification_fail_prompt_builder.table
    - sender: verification_fail_prompt_builder.prompt
      receiver: verification_fail_adapter.prompt
    - sender: verification_fail_adapter.output
      receiver: verification_fail_answer_builder.replies
    - sender: client_database.dataframe
      receiver: verification_router.dataframe
    - sender: verification_router.fail
      receiver: verification_fail_answer_builder.query
    - sender: verification_router.success
      receiver: verification_success_prompt_builder.table
    - sender: verification_success_prompt_builder.prompt
      receiver: verification_success_llm.prompt
    - sender: verification_success_llm.replies
      receiver: verification_success_answer_builder.replies
    - sender: verification_router.success
      receiver: verification_success_answer_builder.query
    - sender: verification_success_answer_builder.answers
      receiver: verification_answer_builder_joiner.answers
    - sender: verification_fail_answer_builder.answers
      receiver: verification_answer_builder_joiner.answers
    - sender: conditional_router.sentiment
      receiver: sentiment_prompt_builder.question
    - sender: sentiment_prompt_builder.prompt
      receiver: sentiment_llm.prompt
    - sender: sentiment_llm.replies
      receiver: sentiment_answer_builder.replies
    - sender: conditional_router.sentiment
      receiver: sentiment_answer_builder.query
    - sender: conditional_router.negotiation
      receiver: negotiation_prompt_builder.question
    - sender: negotiation_prompt_builder.prompt
      receiver: negotiation_llm.prompt
    - sender: negotiation_llm.replies
      receiver: negotiation_answer_builder.replies
    - sender: conditional_router.negotiation
      receiver: negotiation_answer_builder.query
    - sender: conditional_router.escalation
      receiver: escalation_prompt_builder.question
    - sender: escalation_prompt_builder.prompt
      receiver: escalation_llm.prompt
    - sender: escalation_llm.replies
      receiver: escalation_answer_builder.replies
    - sender: conditional_router.escalation
      receiver: escalation_answer_builder.query
    - sender: conditional_router.extract
      receiver: info_extract_prompt_builder.question
    - sender: info_extract_prompt_builder.prompt
      receiver: info_extract_llm.prompt
    - sender: info_extract_llm.replies
      receiver: info_extract_answer_builder.replies
    - sender: conditional_router.escalation
      receiver: info_extract_answer_builder.query
    - sender: conditional_router.payment
      receiver: payment_prompt_builder.question
    - sender: payment_prompt_builder.prompt
      receiver: payment_llm.prompt
    - sender: payment_llm.replies
      receiver: payment_answer_builder.replies
    - sender: conditional_router.payment
      receiver: payment_answer_builder.query
    - sender: verification_answer_builder_joiner.answers
      receiver: answer_joiner.answers
    - sender: initiate_call_answer_builder.answers
      receiver: answer_joiner.answers
    - sender: sentiment_answer_builder.answers
      receiver: answer_joiner.answers
    - sender: negotiation_answer_builder.answers
      receiver: answer_joiner.answers
    - sender: escalation_answer_builder.answers
      receiver: answer_joiner.answers
    - sender: info_extract_answer_builder.answers
      receiver: answer_joiner.answers
    - sender: payment_answer_builder.answers
      receiver: answer_joiner.answers
  max_runs_per_component: 100
  metadata: {}
  inputs:  # Define the inputs for your pipeline
    query:  # These components will receive the query as input
      - "conditional_router.question"
      - "root_prompt_builder.question"
      - "verification_success_prompt_builder.question"
  outputs:  # Defines the output of your pipeline
    answers: "answer_joiner.answers"  # The output of the pipeline is the generated answers
    
indexing_yaml: |
  # If you need help with the YAML format, have a look at https://docs.cloud.deepset.ai/v2.0/docs/create-a-pipeline#create-a-pipeline-using-pipeline-editor.
  # This section defines components that you want to use in your pipelines. Each component must have a name and a type. You can also set the component's parameters here.
  # The name is up to you, you can give your component a friendly name. You then use components' names when specifying the connections in the pipeline.
  # Type is the class path of the component. You can check the type on the component's documentation page.
  # NOTE: The query pipeline does not make use of the document store, as it only makes a call to GPT-4o.
  # However, an indexing pipeline is nonetheless required.
  components:
    file_classifier:
      type: haystack.components.routers.file_type_router.FileTypeRouter
      init_parameters:
        mime_types:
        - text/plain
    text_converter:
      type: haystack.components.converters.txt.TextFileToDocument
      init_parameters:
        encoding: utf-8
    writer:
      type: haystack.components.writers.document_writer.DocumentWriter
      init_parameters:
        document_store:
          type: haystack_integrations.document_stores.opensearch.document_store.OpenSearchDocumentStore
          init_parameters:
            embedding_dim: 768
            similarity: cosine
        policy: OVERWRITE
  connections:  # Defines how the components are connected
  - sender: file_classifier.text/plain
    receiver: text_converter.sources
  - sender: text_converter.documents
    receiver: writer.documents
  inputs:  # Define the inputs for your pipeline
    files: "file_classifier.sources"  # This component will receive the files to index as input