Adding Unstructured Data with Document Narratives

Understanding Document Narratives

A Document Narrative tracks a Document Store.

When files are added or removed from the Document Store:

1. The Narrative filters documents using the selected tag (if one is configured).

2. The selected documents are processed using the chosen parser.

3. The parser breaks each document into smaller sections and prepares them for search.

4. A summary may be generated to capture the key themes of the document.

5. The processed content is added to the Knowledge Store.

This allows assistants to answer questions about document collections such as contracts, research reports, or SEC filings.

The Summarization Prompt is especially important for document narratives. It helps guide the assistant to focus on the types of information you are most likely to ask about.

Creating a Document Narrative

1. Click the Management icon in the application menu.

2. Under AI Assistants, click Knowledge Stores.

3. Click the Document tab.

4. In the row of the Knowledge Store you want to add a Narrative to, click the menu icon.

5. Click + New Narrative.

6. Provide the following information:

   Narrative Name

   Required—Unique identifier for the Narrative.

   Description

   Optional—A description of the Narrative.

   Max Failed Records

   Required—Maximum number of records that can fail processing before the Narrative enters an error state.

   Document Store

   Required—Select the Document Store that will be the source of documents for this Narrative.

   Tag Filter

   Required—Select the tag that identifies which documents from the Document Store are processed and added to this Knowledge Store.

   Summarization Question

   Optional—Provide a prompt to guide the system when generating document summaries and contextualizing document chunks for search. Use this field to emphasize identifying information such as company names, reporting periods, project names, or key entities that users are likely to ask about.

   Avoid including values that vary between sections, such as specific metrics or figures, as this can reduce search accuracy.

   Summarization Chunk Limit

   Required—Specifies the maximum number of document chunks to use when generating the summary. A value of 0 uses all available chunks.

   Parser

   Required—Select the parser to be used to process the documents. Mixed Content is best suited for documents with many visualizations, charts and diagrams that users are likely to query, while Text–Focused is best suited for text-heavy documents.

7. Click Submit.

Next, run the Narrative to process its contents.