GenAI
08-05-2024 | GenAI
Bug fix | Knowledge Preview
The preview of uploaded knowledge didn't show any enters. This has been updated now such that the formatting of the text follows the formatting of the website/file more closely.
07-05-2024 | GenAI
Bug fix | Web Scraper
The web scraper was not able to scrape web pages with unexpected elements in the headings. This has been fixed now.
25-04-2024 | GenAI
Feature release | Interactive Guided Tour
They say a picture is worth a thousand words, but what about an interactive guided tour? We are introducing the first interactive guided tour, which will allow you to get started faster with GenAI (Basic). In less than five minutes, the tour walks you through the most important features: Profile selection, Knowledge base, Behaviour settings, and finally, the Chatbot Test Centre. You can start the tour at any moment in time using the 'Take the tour' button in the left corner of the screen.
18-04-2024 | GenAI
Feature release | Cross Encoder
We have recently enhanced our GenAI engine by integrating a state-of-the-art cross-encoder model, which works in tandem with the existing bi-encoder model. These models are used to find relevant documents that may be used to answer a customers' question. In our new implementation, the added cross-encoder re-evaluates the documents initially retrieved by the bi-encoder to ensure optimal relevance. This update aims to provide more precise and useful answers by ensuring that the most relevant documents are prioritized.
Consider this example. A customer asks: "Where can I find my invoices?" Below is a comparison of how the bi-encoder and the cross-encoder rank relevant documents:
Document | Bi-encoder rank | Cross-encoder Rank |
"You can find your invoices in our monthly email or on our website under 'My Data.'" | 2 | 1 |
"Invoices are issued monthly and are available via email or directly on our website." | 3 | 2 |
"Your invoices contain information on your made costs." | 1 | 3 |
"Any changes to your account will be announced in advance via email." | 4 | 4 |
The example demonstrates that the bi-encoder did not initially prioritize the most relevant documents (1 and 2) correctly. The cross-encoder, by reassessing the bi-encoder's selections, identified and promoted them to the top positions, ensuring a more accurate response to the user's query.
Note: The similarity scores displayed for the top-ranked documents are still derived from the bi-encoder. This means documents with higher bi-encoder similarity scores may be ranked lower if the cross-encoder determines other documents to be more relevant to the query. See an example of this below, where the order of the sources does not correspond to the order of the shown distances, because the shown distances are that of the bi-encoder.
16-04-2024 | GenAI
Feature release | Additional Configuration Options
We've introduced three new fields within the GenAI application across two screens:
- Behaviour:
- Company name: the name of the organisation that the AI will represent.
- Profile Settings:
- API key: required for onboarding GenAI profiles with products like Conversational AI Cloud.
- Product token: required for onboarding channel based integrations with CM.com's Conversational Router.
15-04-2024 | GenAI
Feature release | Behaviour
Introducing a new screen to allow for the configuration of the AI's behaviour under different circumstances, including:
- Tone of voice (moved over from profile settings).
- Base language selection
- Welcome message
- No answer message
- Handover behaviour (enable/disable, questions asked)
11-04-2024 | GenAI
Improvement | Chunking strategy
Various revisions have been done to improve the chunking results from the uploaded knowledge (both for scraped websites and uploaded files):
- It is now possible to read in tables in PDFs
- Chunks will now start with the section title/heading (if present), making them easier to find in the retrieval process when a query is asked
- PDFs consisting of less than 400 words are not chunked, but are instead included in the knowledge as a whole, leading to lower loss of information
- Irreadable chunks are removed from the database (e.g. if they consist of less than 10 words, consist for a large part of spaces, etc.)
- Chunks that are too similar to already-existing chunks are not saved as knowledge. As a result, recurring information (such as headers/footers on websites) will not end up in the knowledge database more than once.
As a result, the database will contain more concise and distinctive chunks, leading to retrieval of more relevant knowledge and thus more relevant answers by the GenAI Engine.
10-04-2024 | GenAI
Feature release | Conversation clustering tracking
Previously every time a synchronisation was performed on conversation history, all already-existing clusters were deleted and overwritten by new clusters. This has been adjusted such that previously-existing clusters remain available, and are automatically mapped to newly-created clusters (if they are similar enough). In the near future, this feature will allow users to enable/disable specific clusters from entering the knowledge database.
Bug Fix | Conversation history storer
The prompts that are used to transform conversation history into summarized question-answer pairs was accidentally hardcoded to the Dutch language instead of the original language, causing non-Dutch conversations to be dropped from the database. This bug is now fixed.
21-03-2024 | GenAI
Released the GenAI application for the CM.com platform. The GenAI app is the core application for customers to manage their AI within the CM.com platform. It provides the following high-level features for customers:
- Self-service profile management:
- A profile is an isolated AI that can be fed distinctive data, and will learn within its own domain separate from other profiles in any customers environment.
- There is no limitation to the amount of profiles a customer can create.
- Knowledge sources, with the initial release we support:
- File uploads (JSON and PDF).
- Web data through our web scraper. Single URL's, or wildcard URLS (urls ending in /*) that crawl multiple pages at once.
- Conversation history, a mechanism that automatically generates new content based on conversations between live agents and end-users.
- Audiences:
- A way for customers to indicate the intended audience of a specific data source, e.g. a customer might only want a particular file to be used for answer suggestion for agents, but not for the conversational AI.
- Synchronisations:
- A synchronisation is a process that fetches the latest data from all connected sources, e.g. configured URL's for the web crawler, and recently uploaded/updated files.
- A fully transparant process that shows what data it extracted from what sources.
- Synchronisation can be configured to run on an automated schedule, e.g. every day or every Sunday.
- An optional web crawler blacklists that tells the system to avoid specific URL's when crawling configured web pages.
- The ability to synchronise individual sources to allow for rapid iterations/testing.
- Test center and diagnostics:
- The ability to talk directly to the AI to test and validate the AI.
- Full diagnostics of AI behaviour, including the reasoning of the AI when formulating a response, and the sources used to formulate that particular response.
Channel integrations:
- WhatsApp Business
- Apple Messages for Business
- Viber for Business
- Facebook Messenger
- Instagram Messaging
- Telegram
- Google's Business Messages
- RCS Business Messaging
- Microsoft Teams
- Slack
All CCaaS integrations made available through marketplace, to name a few:
- Salesforce Service Cloud
- Dynamics Omnichannel
- Quandago
