Tutorial: Building a Summarization System with a Large Language Model

This tutorial teaches you how to build a question answering system that generates answers based on your documents. It uses the PromptNode with a large language model.

  • Level: Beginner
  • Time to complete: 15 minutes
  • Prerequisites:
    • This tutorial assumes a basic knowledge of large language models and retrieval-augmented generation (RAG). If you need more information, have a look at Language Models and Retrieval Augmented Generation (RAG) Question Answering.
    • You must be an Admin to complete this tutorial.
    • This tutorial uses the gpt-3.5-turbo model, so you need an API key from an active OpenAI account.
      If you don't have an account with OpenAI, you can replace this model with an open source one, like Llama2, but bear in mind its performance may not be sufficient.
  • Goal: After completing this tutorial, you will have created a system that can generate summaries of reports on child obesity and food advertising regulations. You will have learned how to use PromptNode with a large language model and a custom prompt.
  • Keywords: PromptNode, summarization, large language models, prompts

Connect Your OpenAI Account

Perform this step if you want to use the gpt-3.5-turbo model by OpenAI. If you're planning to use an open source model, you can skip this step.

You'll be able to use OpenAI models without having to pass the API keys in the pipeline itself.

  1. In deepset Cloud, click your initials in the top right corner and choose Connections.
The personal menu expanded with the Connections option underlined.
  1. Next to OpenAI, click Connect, paste your OpenAI API key, and click Submit.

Result: You're connected to your OpenAI account and can use OpenAI models in your pipelines.

The integrations section with the OpenAI option showing as connected.

Upload Files

First, let's upload the files we want our search system to run on. The files here are a set of reports on the impact of food marketing on child obesity. You can replace this dataset with any other dataset.

  1. Download the .zip file with sample files and unpack it on your computer.

  2. Go to deepset Cloud, make sure you're in the workspace you want to use for this task, and go to _Files.

    The left hand navigation with the workspace name numbered as 1 and the files option numbered as 2
  3. Click Upload Files.

  4. Select all the files you extracted, drop them into the Upload Files window, and click Upload. There should be four files in total.

Result: Your files are in your workspace, and you can see them on the Files page.

The Files page with the four files successfully uploaded.

Create the Pipeline

We'll use an out-of-the-box template as a baseline for our pipeline and we'll adjust it a bit:

  1. In deepset Cloud, go to Pipeline Templates.

  2. Click Basic QA, find Retrieval Augmented Generation Question Answering GPT-3.5, and choose Use Template.

    The templates page with template cards showing. On the left, in the navigation, the Basic QA option is selected and marked as step 1. Then, a template called Retrieval Augmented Generation Question Answering gpt-3.5 is highlighted and an option Use Template is marked as step 2.
  3. Type summarization as the pipeline name and click Create Pipeline. You're redirected to the Pipelines page. You can find your pipeline in the All tab.

  4. Click the More Actions button next to your pipeline and choose Edit.

    The Pipelines page with the more actions button next to the summarization pipeline expanded and the edit option highlighted.t
  5. Copy this pipeline configuration and paste it into the Code Editor:

      - name: DocumentStore
        type: DeepsetCloudDocumentStore
          embedding_dim: 768
          similarity: cosine
      - name: BM25Retriever # The keyword-based retriever
        type: BM25Retriever
          document_store: DocumentStore
          top_k: 10 # The number of results to return
      - name: EmbeddingRetriever # Selects the most relevant documents from the document store
        type: EmbeddingRetriever # Uses a Transformer model to encode the document and the query
          document_store: DocumentStore
          embedding_model: intfloat/e5-base-v2 # Model optimized for semantic search. It has been trained on 215M (question, answer) pairs from diverse sources.
          model_format: sentence_transformers
          top_k: 10 # The number of results to return
      - name: JoinResults # Joins the results from both retrievers
        type: JoinDocuments
          join_mode: concatenate # Combines documents from multiple retrievers
      - name: Reranker # Uses a cross-encoder model to rerank the documents returned by the two retrievers
        type: SentenceTransformersRanker
          model_name_or_path: intfloat/simlm-msmarco-reranker # Fast model optimized for reranking
          top_k: 1 # The number of results to return
          batch_size: 20  # Try to keep this number equal or larger to the sum of the top_k of the two retrievers so all docs are processed at once
          model_kwargs:  # Additional keyword arguments for the model
            torch_dtype: torch.float16
      - name: summarization
        type: PromptTemplate
            type: AnswerParser
          prompt: deepset/summarization
      - name: PromptNode
        type: PromptNode
          default_prompt_template: summarization
          max_length: 400 # The maximum number of tokens the generated answer can have
          model_kwargs: # Specifies additional model settings
            temperature: 0 # Lower temperature works best for fact-based qa
          model_name_or_path: gpt-3.5-turbo
          top_k: 1
      - name: ReferencePredictor # Finds and displays references to the generated answers from the retrieved documents
        type: ReferencePredictor
          model_name_or_path: cross-encoder/ms-marco-MiniLM-L-6-v2
          verifiability_model_name_or_path: tstadel/answer-classification-setfit-v2-binary
          language: en
          use_split_rules: True # Uses additional rules for better splitting
          extend_abbreviations: True # Extends abbreviations handled with a curated list
      - name: FileTypeClassifier # Routes files based on their extension to appropriate converters, by default txt, pdf, md, docx, html
        type: FileTypeClassifier
      - name: TextConverter # Converts files into documents
        type: TextConverter
      - name: PDFConverter # Converts PDFs into documents
        type: PDFToTextConverter
      - name: Preprocessor # Splits documents into smaller ones and cleans them up
        type: PreProcessor
          # With a vector-based retriever, it's good to split your documents into smaller ones
          split_by: word # The unit by which you want to split the documents
          split_length: 250 # The max number of words in a document
          split_overlap: 20 # Enables the sliding window approach
          language: en
          split_respect_sentence_boundary: True # Retains complete sentences in split documents
    # Here you define how the nodes are organized in the pipelines
    # For each node, specify its input
      - name: query
          - name: BM25Retriever
            inputs: [Query]
          - name: EmbeddingRetriever
            inputs: [Query]
          - name: JoinResults
            inputs: [BM25Retriever, EmbeddingRetriever]
          - name: Reranker
            inputs: [JoinResults]
          - name: PromptNode
            inputs: [Reranker]
          - name: ReferencePredictor
            inputs: [PromptNode]
      - name: indexing
        # Depending on the file type, we use a Text or PDF converter
          - name: FileTypeClassifier
            inputs: [File]
          - name: TextConverter
            inputs: [FileTypeClassifier.output_1] # Ensures that this converter receives txt files
          - name: PDFConverter
            inputs: [FileTypeClassifier.output_2] # Ensures that this converter receives PDFs
          - name: Preprocessor
            inputs: [TextConverter, PDFConverter]
          - name: EmbeddingRetriever
            inputs: [Preprocessor]
          - name: DocumentStore
            inputs: [EmbeddingRetriever]
    1. Here's an explanation of the changes:
      1. In line 37, we changed the top_k value of the SentenceTransformersRanker component with 1 to make sure the Ranker returns one best document for the summary.
      2. In line 41, we changed the prompt template name to summarization and replaced the prompt text with a link to a ready-made summarization template (deepset/summarization).
      3. In line 50, we pointed PromptNode to use the updated summarization prompt.
      4. Line 54 is where you can change the LLM PromptNode uses.
      5. In line 55, we added the top_k: 1 parameter to make sure PromptNode returns only one summary.
  6. Save your pipeline.

  7. At the top of the Designer, click Deploy and wait until your pipeline is deployed and indexed. Indexing may take a couple of minutes.

Result: You created a pipeline summarizing documents using a large language model. The pipeline status is indexed, which means it's ready for use. Your pipeline is at the development service level. We recommend you test it before setting it to the production service level.

Test the Pipeline

Now it's time to see how your pipeline is doing. Let's run a search with it.

  1. In the navigation, click Playground.

  2. Make sure the summarization pipeline is selected.

  3. Type the query: summarize the report on advertising food to children.
    Here's what the pipeline returns:
    The Search page with a summary the pipeline returned as an answer to the query.

Result: Congratulations! You just created a summarization pipeline that uses a large language model to generate summaries of documents.

What's Next

Your pipeline is now a development pipeline. Once it's ready for production, change its service level to Production. You can do this on the Pipeline Details page shown after clicking a pipeline name. To learn more, see Pipeline Service Levels.