Skip to content

⭐ Core Concepts

Glossary

General and 🦑TruLens-specific concepts.

  • Agent. A Component of an Application or the entirety of an application that provides a natural language interface to some set of capabilities typically incorporating Tools to invoke or query local or remote services, while maintaining its state via Memory. The user of an agent may be a human, a tool, or another agent. See also Multi Agent System.

  • Application or App. An "application" that is tracked by 🦑TruLens. Abstract definition of this tracking corresponds to App. We offer special support for LangChain via TruChain, LlamaIndex via TruLlama, and NeMo Guardrails via TruRails Applications as well as custom apps via TruBasicApp or [TruApp][trulens.apps.app.TruApp], and apps that already come with Traces via TruVirtual.

  • Chain. A LangChain App.

  • Chain of Thought. The use of an Agent to deconstruct its tasks and to structure, analyze, and refine its Completions.

  • Completion, Generation. The process or result of LLM responding to some Prompt.

  • Component. Part of an Application giving it some capability. Common components include:

  • Retriever

  • Memory

  • Tool

  • Agent

  • Prompt Template

  • LLM

  • Embedding. A real vector representation of some piece of text. Can be used to find related pieces of text in a Retrieval.

  • Eval, Evals, Evaluation. Process or result of method that scores the outputs or aspects of a Trace. In 🦑TruLens, our scores are real numbers between 0 and 1.

  • Feedback. See Evaluation.

  • Feedback Function. A method that implements an Evaluation. This corresponds to Feedback.

  • Fine-tuning. The process of training an already pre-trained model on additional data. While the initial training of a Large Language Model is resource intensive (read "large"), the subsequent fine-tuning may not be and can improve the performance of the LLM on data that sufficiently deviates or specializes its original training data. Fine-tuning aims to preserve the generality of the original and transfer of its capabilities to specialized tasks. Examples include fine-tuning on:

  • financial articles

  • medical notes

  • synthetic languages (programming or otherwise)

While fine-tuning generally requires access to the original model parameters, some model providers give users the ability to fine-tune through their remote APIs.

  • Generation. See Completion.

  • Human Feedback. A feedback that is provided by a human, e.g. a thumbs up/down in response to a Completion.

  • In-Context Learning. The use of examples in an Instruction Prompt to help an LLM generate intended Completions. See also Shot.

  • Instruction Prompt, System Prompt. A part of a Prompt given to an LLM to complete that contains instructions describing the task that the Completion should solve. Sometimes such prompts include examples of correct or intended completions (see Shots). A prompt that does not include examples is said to be Zero Shot.

  • Language Model. A model whose task is to model text distributions typically in the form of predicting token distributions for text that follows the given prefix. Proprietary models usually do not give users access to token distributions and instead Complete a piece of input text via multiple token predictions and methods such as beam search.

  • LLM, Large Language Model (see Language Model). The Component of an Application that performs Completion. LLMs are usually trained on a large amount of text across multiple natural and synthetic languages. They are also trained to follow instructions provided in their Instruction Prompt. This makes them general in that they can be applied to many structured or unstructured tasks and even tasks which they have not seen in their training data (See Instruction Prompt, In-Context Learning). LLMs can be further improved for rare/specialized settings using Fine-Tuning.

  • Memory. The state maintained by an Application or an Agent indicating anything relevant to continuing, refining, or guiding it towards its goals. Memory is provided as Context in Prompts and is updated when new relevant context is processed, be it a user prompt or the results of the invocation of some Tool. As Memory is included in Prompts, it can be a natural language description of the state of the app/agent. To limit the size of memory, Summarization is often used.

  • Multi-Agent System. The use of multiple Agents incentivized to interact with each other to implement some capability. While the term predates LLMs, the convenience of the common natural language interface makes the approach much easier to implement.

  • Prompt. The text that an LLM completes during Completion. In chat applications. See also Instruction Prompt, Prompt Template.

  • Prompt Template. A piece of text with placeholders to be filled in in order to build a Prompt for a given task. A Prompt Template will typically include the Instruction Prompt with placeholders for things like Context, Memory, or Application configuration parameters.

  • Provider. A system that provides the ability to execute models, either LLMs or classification models. In 🦑TruLens, Feedback Functions make use of Providers to invoke models for Evaluation.

  • RAG, Retrieval Augmented Generation. A common organization of Applications that combine a Retrieval with an LLM to produce Completions that incorporate information that an LLM alone may not be aware of.

  • RAG Triad (🦑TruLens-specific concept). A combination of three Feedback Functions meant to Evaluate Retrieval steps in Applications.

  • Record. A "record" of the execution of a single execution of an app. Single execution means invocation of some top-level app method. Corresponds to Record

    Note

    This will be renamed to Trace in the future.

  • Retrieval, Retriever. The process or result (or the Component that performs this) of looking up pieces of text relevant to a Prompt to provide as Context to an LLM. Typically this is done using Embedding representations.

  • Selector (🦑TruLens-specific concept). A specification of the source of data from a Trace to use as inputs to a Feedback Function. This corresponds to Lens and utilities Select.

  • Shot, Zero Shot, Few Shot, <Quantity>-Shot. Zero Shot describes prompts that do not have any examples and only offer a natural language description of the task to be solved, while <Quantity>-Shot indicate some <Quantity> of examples are provided. The "shot" terminology predates instruction-based LLMs where techniques then used other information to handle unseen classes such as label descriptions in the seen/trained data. In-context Learning is the recent term that describes the use of examples in Instruction Prompts.

  • Span. Some unit of work logged as part of a record. Corresponds to current 🦑RecordAppCallMethod.

  • Summarization. The task of condensing some natural language text into a smaller bit of natural language text that preserves the most important parts of the text. This can be targeted towards humans or otherwise. It can also be used to maintain concise Memory in an LLM Application or Agent. Summarization can be performed by an LLM using a specific Instruction Prompt.

  • Tool. A piece of functionality that can be invoked by an Application or Agent. This commonly includes interfaces to services such as search (generic search via Google or more specific like IMDB for movies). Tools may also perform actions such as submitting comments to GitHub issues. A Tool may also encapsulate an interface to an Agent for use as a component in a larger Application.

  • Trace. See Record.