Building user journeys as declarative trees within a virtual assistant requires assumptions to be made about the user query and the optimal path. If there are many decision points and the tree consists of many forks the number of assumptions increases exponentially down the tree leading to inefficiencies and a suboptimal design. To address this inefficiency, one approach is to use a language model to reason over available tools (or APIs) that can be called to augment the response to the query. This collapses the tree and replaces it with a language model that can be guided through a policy or rules expressed in natural language and supplied to the model in a prompt.
Dec 6, 2023
Fine-tuning of a large language model (LLM) can be peformed using QLoRA (Quantized Low Rank Adapters) and PEFT (Parameter-Efficient Fine-Tuning) techniques. PEFT (Parameter-Efficient Fine-Tuning): PEFT is a technique for fine-tuning large language models with a small number of additional parameters, known as adapters, while freezing the original model parameters. It allows for efficient fine-tuning of language models, reducing the memory footprint and computational requirements. PEFT enables the injection of niche expertise into a foundation model without catastrophic forgetting, preserving the original model’s performance. LoRA (Low Rank Adapters):
Dec 1, 2023
Retrieval Augmented Generation (RAG), which utilises a LLM, makes it relatively straightfoward to surface information through a conversational assistant. This is potentially transformative for HR & talent management and customer care use cases where information contained in policies, guidelines, handbooks and other unstructured natural language formats can be made more accessible and conveniently queried through an assistant’s natural language interface. Here I share an architecture that extends a conversational assistant with RAG, routing searches to collections mapped to a user and intent.
Nov 4, 2023
Large Language Models (LLMs) can be used to automate testing of virtual assistants. One approach is to use the LLM to generate the queries and responses of the human user to automate the test of a journey, end to end. Here I share a conceptual data pipeline view of such a system. The key ideas are:
Nov 1, 2023
View the post here: https://jamesdhope.medium.com/graph-driven-llm-assisted-virtual-assistant-architecture-c1e4857a7040.
Oct 2, 2023