The client has accumulated digital assets related to its business operations, including product details, technical articles and customer interaction logs, stored in a variety of formats. The goal was to make the knowledge retrievable both using keyword search and a natural-language interface organized in the manner of a chatbot.
We started out building a chat-based service to optimize current business processes, then the team realized the technology could help other businesses, and initiated building a new AI-powered product for sales teams.
Assessment
- Leadership team is enthused about making internal digital assets available to all employees in a chat-style user interface akin to ChatGPT or Anthropic Claude.
- In-house software team is highly motivated to learn how to apply emerging NLP methodologies in information retrieval.
- Client is interested in exploring the productization of a private search engine with a chat interface to other businesses for sales productivity.
Our assessment results are summarized below:
AI Business Problem and Value
The client has accumulated detailed content over time that capture business processes, product details, technical articles and customer interaction logs. These are stored in a multiple formats in a loosely organized data store that includes the following:
- Structured documents organized in a SQL database
- Partially structured documents organized in Elasticsearch
- Web pages written over time for use in the customer-facing website
- PDF and Word documents written by different people over time in their own individual writing style. These were placed in a loosely structured network file system with file names that did not convey enough context of the content of each file
An important objective was to create a user-friendly chat-style interface to allow internal users to query and retrieve relevant information from the organizations’ proprietary knowledge base. This would help unlock the information latent in the disparate collection of documents without requiring technical skills such as SQL queries or having to sift manually through a lot of documents to locate the useful information.
AI Use cases
We identified the following potential AI use cases for internal users
- Natural Language Query: Enable employees to quickly find relevant information from various data sources such as manuals, reports, presentations, and project documentation
- Keyword Query: In addition to natural language query the employees would like to be able to use conventional keyword queries to locate relevant information
- Summarization: The results should enable an optional summary of the results
- Citation: Source documents should be available for deeper study
The capabilities were expected to be widely used within the company. E..g, they would enable the sales and marketing team to locate and access past marketing collateral, product brochures, case studies, and customer success stories from past activities.
Modeling Approach
We reviewed the applicability of both traditional AI and Gen AI techniques as appropriate and recommended extraction of metadata from each document and storing in the data store to support retrieval.
Traditional AI
We used a variety of analytical AI/ML techniques:
- Extraction of keywords and business-specific entities
- Classification of each document into apt categories, e.g., Technical paper, Business brief, Product marketing, Finance, etc.
- Topic extraction, i.e., a brief descriptor of what is covered in each document
- Vectorization, i.e., storing a numerical representation of each document for use in vector search
Generative AI
A publicly available LLM was leveraged to extract the following from each document:
- Summarization of the content of each document - written in a consistent manner
- Extraction of text content from PDF, Word, and HTML documents
AI Data Strategy
We recommended storing all documents in a common document-oriented data store. Each document was enriched with content extracted by AI models as described above with the following:
- Data indexing: Index all data sources into a centralized document-oriented data store
- Knowledge Graph: i.e., hierarchical representation of the concepts represented within the document store was made available as an API
- Data Augmentation: Public data sets that aligned with the customer’s business domain were incorporated to enrich the query results
User Interface
- Chat-style interface with a conversational experience allowing users to type in their queries and receive relevant search results. While most queries were expected to be searches for documents, the interface supported a continuation of interaction based on the context of the user’s interaction.
- Autocomplete and Suggested Queries: The user was shown autocomplete and query suggestions as they typed in their query
- Result formatting: The search results optionally showed a summarized answer to the query and a list of reference to the source documents
- Ability to drill down and view the reference documents
- Log in using users' company credentials. Runtime LLM charges incurred owing to a user query were tracked per user - without saving details of each query itself.
Results
The client accepted the recommendations and was able to build a proof-of-concept in a few weeks and an internal release to employees in a few months without hiring more resources. A seasoned data scientist helped with the system design and supervised the work of the internal team during the development phase.
Future work is required to build this out as a new standalone AI-powered product for sale to new customers.
