May 2023 Summaries

The text discusses the shift in AI design from natural language understanding (NLU) to natural language generation (NLG), with a focus on controlling generated output, persona design, knowledge management, bias, and prompt chaining. As LLMs become more prevalent, designers will need to adapt their approach to crafting flexible prompts that can generate desired outputs, designing personas that capture the human touch, managing knowledge bases to improve model accuracy, mitigating bias and oversight issues, and developing skills like prompt chaining to create repeatable outcomes. The future of AI-generated conversations is expected to be shaped by these changes, with a focus on interpretability, privacy, and oversight.

The author emphasizes the importance of understanding your specific situation before trying to adopt others' practices or "best practices" in AI-powered assistant design. They suggest that without clear answers to questions about your company's size, customer base, goals, and desired properties for your assistant, it's unlikely that someone will be able to create a work-fluencing LinkedIn post that applies perfectly to you. The author advocates for secret shopping the competition by trying out various assistants and services, including those from competitors and within your own organization, to gather novel experiences and feedback. They propose a discussion process where everyone on the team collects and shares these experiences, allows personal preferences to bleed into choices, holds a total teardown of ideas, prioritizes winning concepts, and tests them further. The author believes that this cycle is essential for creating effective AI-powered assistant designs, but acknowledges that "good" and "success" are subjective and ephemeral, and that the key to success lies in having good conversations.

AI Builder is a collaborative and secure enterprise platform that allows users to create instant assistants using Voiceflow's platform. With Knowledge Base, users can upload a PDF and build a GPT-powered chatbot. Generative Tasks enable the creation of response variations, utterances, entities, and synonyms, while AI Freestyle allows for dynamic conversation flow. Response AI Step dynamically generates user responses based on prompts, allowing for faster and easier design. LLM functions enable variable setting in real-time, unlocking use cases like branch logic, classifying user utterances, and validating data. The platform also offers a conversation memory feature to provide context and automatically recall previous conversations. Users can get started with AI Builder by signing up for Voiceflow, creating a new assistant, and selecting the "Build AI Assistant" project type. The platform provides templates for inspiration, including customer support automation, travel assistance, sales agent, character role-playing, and conversational knowledge bases.

To integrate Claude LLM into a Voiceflow assistant, first sign up for early access and gain access to the Anthropic Developer Console, then generate an API key. From there, follow steps to create an integration by logging in to your Voiceflow account, choosing the project, reaching for the API step, entering the API endpoint URL, setting the API method and headers, including the API key, and providing the required JSON payload. You can test the assistant while prototyping but note that commercial deployment requires a services agreement with Anthropic and evaluation API keys are provided under non-commercial terms.

You'll need Node.js 18+ installed on your system, as well as Docker Compose, to run this code. To get started, you should copy and set up the `.env` file with required environment variables. Running the provided command will create several containers, including Redis, Unstructured, OpenSearch, and OpenSearch-dashboards. The OpenSearch dashboard can be accessed at http://localhost:5601. There are multiple API endpoints available, such as `POST /api/add`, `GET /api/live`, and `DELETE /api/collection`, which provide various tasks like adding content to OpenSearch, getting responses using a live webpage or vector store, clearing Redis cache, and deleting collections. You can use the `/api/live` endpoint to get responses without vectorizing the content, and once you've added content to OpenSearch, you can use the `/api/question` endpoint to get answers based on your knowledge base, which can be integrated into Voiceflow Assistants for answering user queries.