Skip to main content
Ghost Weights™: Systematic assessments that measure model performance against specific criteria or benchmarks. Evaluations help you understand accuracy, consistency, and quality across different types of queries. The 4Minds platform supports automated testing, human review, and model judging to ensure your AI meets your organization’s standards before and during deployment.Ground Truth: The verified correct answer used to evaluate model performance in question-and-answer pairs. In 4Minds, ground truths are generated directly from your uploaded data rather than generic public datasets. This ensures evaluations test how well your model performs on your actual business context and proprietary knowledge, not just on industry-standard benchmark datasets like MMLU (Massive Multitask Language Understanding) or SQuAD (Stanford Question Answering Dataset).Synthesis Graph™: 4Minds’ technology that automatically organizes your documents into intelligent, hierarchical knowledge structures that function as attention layers within the model architecture. Unlike approaches that only modify weights, Synthesis Graph™ works at the transformer level to understand how your information connects and routes queries efficiently across large-scale data, surfacing insights and relationships you might otherwise miss. It stays current as your information grows and changes.Reflex Router™: 4Minds’ intelligent routing system that automatically determines the best way to answer each query. You get the most accurate responses without configuring anything. Reflex Router™ handles the complexity behind the scenes, ensuring fast answers for simple questions and deep insights for complex ones.Inline Tuning™: A 4Minds’ platform feature enabling continuous model adaptation during active use. Your models get smarter as you use them, learning from interactions without requiring manual updates or service interruptions. Your models improve continuously without taking your system offline for updates.Note: Terms marked with ™ are proprietary 4Minds technologies.
Evals (Evaluations): Systematic assessments that measure model performance against specific criteria or benchmarks. Evaluations help you understand accuracy, consistency, and quality across different types of queries. The 4Minds platform supports automated testing, human review, and model judging to ensure your AI meets your organization’s standards before and during deployment.Prompt: The input text or question you provide to a model. Effective prompts help the model understand what you’re asking and generate better responses. The 4Minds platform uses your prompts along with context from your datasets to deliver accurate, relevant answers.
Base Model: The underlying foundation model selected for personalization within the 4Minds platform. Examples include Gemma-12B, Nemotron-70B, and Qwen-32B. Choosing the right base model affects speed, accuracy, and capability—4Minds offers multiple options so you can balance performance with your specific needs. The base model provides core language capabilities before Ghost Weights™ adaptation.Context Window: The maximum amount of text (measured in tokens) a model can process in a single interaction. Context windows determine how much conversation history, document content, or retrieved information can inform each response. Larger context windows enable more comprehensive analysis and better understanding of complex queries. Common examples include 8K, 32K, or 128K tokens.Fine-tuning: The process of adapting a pre-trained model to perform specific tasks or match particular domains by training on targeted datasets. Traditional fine-tuning updates all or most model parameters, making it resource-intensive and requiring service downtime.Inference: The process of generating outputs from a trained model in production. Inference occurs when a deployed model processes queries and produces responses for actual use cases. Faster inference means quicker answers and better user experience.Knowledge Graph: A structured representation of information as entities (nodes) and relationships (edges) that maps how concepts connect. Traditional knowledge graphs are often manually curated or require rigid schemas, making them static and labor-intensive to maintain. They typically struggle to scale beyond limited domains and can’t efficiently route queries across large information spaces, often requiring manual configuration to determine which parts of the graph are relevant for specific questions.LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning technique that updates a subset of model weights through low-rank matrix decomposition. LoRA operates at the weight level, making it more efficient than full fine-tuning but still requiring service downtime for model updates and lacking support for continuous adaptation. It doesn’t address how models process and route information at the transformer layer.Multimodal: AI systems capable of processing and generating multiple types of content beyond text, such as images, audio, video, and documents. Multimodal models can understand relationships across different content types—for example, answering questions about images or extracting information from visual documents. Traditional multimodal implementations often struggle with maintaining context across modalities and require separate processing pipelines for each content type, adding complexity to deployment and maintenance.Parameters: The learned weights within a neural network that determine model behavior. Parameter count (e.g., 12B for 12 billion parameters) generally indicates model capability, though architecture efficiency varies. 4Minds’ Ghost Weights™ delivers better performance by training smarter—adapting less than 5% of parameters—rather than requiring massive parameter counts.Quantization: A technique that reduces model size and speeds up inference by converting high-precision numerical values to lower-precision formats. Quantization makes models more efficient and cost-effective to run, enabling faster responses with less computational resources. The tradeoff is that aggressive quantization can impact model accuracy. Modern quantization methods like AWQ (Activation-Aware Weight Quantization) minimize accuracy loss while maximizing efficiency gains.RAG (Retrieval-Augmented Generation): An architecture pattern that retrieves relevant information from documents before generating responses, ensuring answers are grounded in actual content rather than the model’s general knowledge. Traditional RAG implementations rely primarily on vector similarity matching, which excels at finding documents with matching keywords but struggles with contextual understanding, multi-hop reasoning, and discovering relationships between concepts. This can lead to missed insights when answers require connecting information across multiple sources or understanding how ideas relate.Token: The fundamental unit of text processing in language models. A token typically represents a word, part of a word, or punctuation mark. Token counts determine context window limits and affect processing speed and cost.Training: The process of teaching a model to perform specific tasks by exposing it to examples and data. Training can range from initial model development to ongoing adaptation. In the 4Minds platform, Ghost Weights™ and Inline Tuning™ enable efficient, continuous training without disrupting your operations. In 4Minds, training a model involves three key steps: creating a model, feeding it a dataset, and iteratively refining it through testing in the Playground where you can add more data as you discover knowledge gaps.Vector Search: A search technique that finds information based on semantic similarity rather than exact keyword matches. Vector search converts text into numerical representations (embeddings) and identifies content with similar meanings. While effective for basic retrieval, vector search alone misses contextual relationships between information, which is why the 4Minds platform combines it with Synthesis Graph™ for more intelligent results.
Can’t find a term? Contact our support team or suggest additions to this glossary.