diff --git a/docs/evaluation/concepts/index.mdx b/docs/evaluation/concepts/index.mdx index 36000d7a..2c41ca4c 100644 --- a/docs/evaluation/concepts/index.mdx +++ b/docs/evaluation/concepts/index.mdx @@ -1,365 +1,173 @@ -# Evaluation concepts +Evaluation Concepts +=================== -The pace of AI application development is often limited by high-quality evaluations. -Evaluations are methods designed to assess the performance and capabilities of AI applications. +High-quality evaluations are key to creating, refining, and validating AI applications. In LangSmith, you’ll find the tools you need to structure these evaluations so that you can iterate efficiently, confirm that changes to your application improve performance, and ensure that your system continues to work as intended. -Good evaluations make it easy to iteratively improve prompts, select models, test architectures, and ensure that deployed applications continue to perform as expected. -LangSmith makes building high-quality evaluations easy. +This guide explores LangSmith’s evaluation framework and core concepts, including: -This guide explains the key concepts behind the LangSmith evaluation framework and evaluations for AI applications more broadly. -The core components of LangSmith evaluations are: +- **Datasets**, which hold test examples for your application’s inputs (and, optionally, reference outputs). +- **Evaluators**, which assess how well your application’s outputs align with the desired criteria. -- [**Datasets**:](/evaluation/concepts#datasets) Collections of test inputs and, optionally, reference outputs for your applications. -- [**Evaluators**](/evaluation/concepts#evaluators): Functions for scoring the outputs generated by applications given dataset inputs. +Datasets +-------- -## Datasets - -A dataset contains a collection of examples used for evaluating an application. +A dataset is a curated set of test examples. Each example can include inputs (the data you feed into your application), optional reference outputs (the “gold-standard” or target answers), and any metadata you find helpful. ![Dataset](./static/dataset_concept.png) ### Examples -Each example consists of: - -- **Inputs**: a dictionary of input variables to pass to your application. -- **Reference outputs** (optional): a dictionary of reference outputs. These do not get passed to your application, they are only used in evaluators. -- **Metadata** (optional): a dictionary of additional information that can be used to create filtered views of a dataset. +Each example corresponds to a single test case. In most scenarios, an example has three components. First, it has one or more inputs provided as a dictionary. Next, there can be reference outputs (if you have a known response to compare against). Finally, you can attach metadata in a dictionary format to store notes or tags, making it easy to slice, filter, or categorize examples later. ![Example](./static/example_concept.png) -### Dataset curation - -There are various ways to build datasets for evaluation, including: - -#### Manually curated examples - -This is how we typically recommend people get started creating datasets. -From building your application, you probably have some idea of what types of inputs you expect your application to be able to handle, -and what "good" responses may be. -You probably want to cover a few different common edge cases or situations you can imagine. -Even 10-20 high-quality, manually-curated examples can go a long way. - -#### Historical traces - -Once you have an application in production, you start getting valuable information: how are users actually using it? -These real-world runs make for great examples because they're, well, the most realistic! - -If you're getting a lot of traffic, how can you determine which runs are valuable to add to a dataset? -There are a few techniques you can use: - -- **User feedback**: If possible - try to collect end user feedback. You can then see which datapoints got negative feedback. - That is super valuable! These are spots where your application did not perform well. - You should add these to your dataset to test against in the future. -- **Heuristics**: You can also use other heuristics to identify "interesting" datapoints. For example, runs that took a long time to complete could be interesting to look at and add to a dataset. -- **LLM feedback**: You can use another LLM to detect noteworthy runs. For example, you could use an LLM to label chatbot conversations where the user had to rephrase their question or correct the model in some way, indicating the chatbot did not initially respond correctly. +### Dataset Curation -#### Synthetic data +When constructing datasets, there are a few common ways to ensure they match real-world use cases: -Once you have a few examples, you can try to artificially generate some more. -It's generally advised to have a few good hand-crafted examples before this, as this synthetic data will often resemble them in some way. -This can be a useful way to get a lot of datapoints, quickly. +- **Manually Curated Examples**. This includes handpicking representative tasks and responses that illustrate normal usage and tricky edge cases. Even a small selection of 10 to 20 examples can yield substantial insights. +- **Historical Traces**. If your system is already in production, gather actual production runs, including examples flagged as problematic by users or system logs. Filtering based on complaints, repeating questions, anomaly detection, or LLM-as-judge feedback can provide a realistic snapshot of real-world usage. +- **Synthetic Generation**. A language model can help you automatically generate new test scenarios, which is especially efficient if you have a baseline set of high-quality examples to guide it. ### Splits -When setting up your evaluation, you may want to partition your dataset into different splits. For example, you might use a smaller split for many rapid and cheap iterations and a larger split for your final evaluation. In addition, splits can be important for the interpretability of your experiments. For example, if you have a RAG application, you may want your dataset splits to focus on different types of questions (e.g., factual, opinion, etc) and to evaluate your application on each split separately. +To organize your dataset, LangSmith allows you to create one or more splits. Splits let you isolate subsets of data for targeted experiments. For instance, you might keep a small “dev” split for rapid iterating and a larger “test” split for comprehensive performance checks. In a retrieval-augmented generation (RAG) system, for example, you could divide data between factual vs. opinion-oriented queries. -Learn how to [create and manage dataset splits](/evaluation/how_to_guides/manage_datasets_in_application#create-and-manage-dataset-splits). +Read more about creating and managing splits [here](/evaluation/how_to_guides/manage_datasets_in_application#create-and-manage-dataset-splits). ### Versions -Datasets are [versioned](/evaluation/how_to_guides/version_datasets) such that every time you add, update, or delete examples in your dataset, a new version of the dataset is created. -This makes it easy to inspect and revert changes to your dataset in case you make a mistake. -You can also [tag versions](/evaluation/how_to_guides/version_datasets#tag-a-version) of your dataset to give them a more human-readable name. -This can be useful for marking important milestones in your dataset's history. - -You can run evaluations on specific versions of a dataset. This can be useful when running evaluations in CI, to make sure that a dataset update doesn't accidentally break your CI pipelines. - -## Evaluators - -Evaluators are functions that score how well your application performs on a particular example. - -#### Evaluator inputs - -Evaluators receive these inputs: - -- [Example](/evaluation/concepts#examples): The example(s) from your [Dataset](/evaluation/concepts#datasets). Contains inputs, (reference) outputs, and metadata. -- [Run](/observability/concepts#runs): The actual outputs and intermediate steps (child runs) from passing the example inputs to the application. - -#### Evaluator outputs - -An evaluator returns one or more metrics. These should be returned as a dictionary or list of dictionaries of the form: - -- `key`: The name of the metric. -- `score` | `value`: The value of the metric. Use `score` if it's a numerical metric and `value` if it's categorical. -- `comment` (optional): The reasoning or additional string information justifying the score. +Every time your dataset changes—if you add new examples, edit existing ones, or remove any entries—LangSmith creates a new version automatically. This ensures you can always revert or revisit earlier states if needed. You can label versions with meaningful tags to mark particular stages of your dataset, and you can run evaluations on any specific version for consistent comparisons over time. -#### Defining evaluators +Further details on dataset versioning are provided [here](/evaluation/how_to_guides/version_datasets). -There are a number of ways to define and run evaluators: +Evaluators +---------- -- **Custom code**: Define [custom evaluators](/evaluation/how_to_guides/custom_evaluator) as Python or TypeScript functions and run them client-side using the SDKs or server-side via the UI. -- **Built-in evaluators**: LangSmith has a number of built-in evaluators that you can configure and run via the UI. +Evaluators assign metrics or grades to your application’s outputs, making it easier to see how well those outputs meet your desired standards. -You can run evaluators using the LangSmith SDK ([Python](https://docs.smith.langchain.com/reference/python) and TypeScript), via the [Prompt Playground](../../prompt_engineering/concepts#prompt-playground), or by configuring [Rules](../../observability/how_to_guides/monitoring/rules) to automatically run them on particular tracing projects or datasets. +### Techniques -#### Evaluation techniques +Below are common strategies for evaluating outputs from large language models (LLMs): -There are a few high-level approaches to LLM evaluation: +- **Human Review**. You and your team can manually assess outputs for correctness and user satisfaction. Use LangSmith Annotation Queues for a structured workflow, including permissions, guidelines, and progress tracking. +- **Heuristic Checking**. Basic rule-based evaluators help detect issues such as empty responses, excessive length, or missing essential keywords. +- **LLM-as-Judge**. A language model can serve as the evaluator, typically via a dedicated prompt that checks correctness, helpfulness, or style. This method works with or without reference outputs. +- **Pairwise Comparisons**. When deciding between two application versions, it can be simpler to ask, “Which output is better?” rather than to assign absolute scores, especially in creative tasks like summarization. -### Human +### Defining Evaluators -Human evaluation is [often a great starting point for evaluation](https://hamel.dev/blog/posts/evals/#looking-at-your-traces). LangSmith makes it easy to review your LLM application outputs as well as the traces (all intermediate steps). +You can use LangSmith’s built-in evaluators or build your own in Python or TypeScript. You can then run these evaluators through the LangSmith SDK, the Prompt Playground (inside LangSmith), or in any automation pipeline you set up. +#### Evaluator Inputs -LangSmith's [annotation queues](/evaluation/concepts#annotation-queues) make it easy to get human feedback on your application's outputs. +An evaluator has access to both the example (input data and optional reference outputs) and the run (the live output from your application). Because each run often includes details like the final answer or any intermediate steps (e.g., tool calls), evaluators can capture nuanced performance metrics. -### Heuristic +#### Evaluator Outputs -Heuristic evaluators are deterministic, rule-based functions. These are good for simple checks like making sure that a chatbot's response isn't empty, that a snippet of generated code can be compiled, or that a classification is exactly correct. +Evaluators usually produce responses in the form of dictionaries (or lists of dictionaries). Each entry typically contains: -### LLM-as-judge +- A “key” or name for the metric. +- A “score” or “value” (numeric or categorical). +- An optional “comment” to explain how or why the score was assigned. -LLM-as-judge evaluators use LLMs to score the application's output. To use them, you typically encode the grading rules / criteria in the LLM prompt. They can be reference-free (e.g., check if system output contains offensive content or adheres to specific criteria). Or, they can compare task output to a reference output (e.g., check if the output is factually accurate relative to the reference). +Experiment +---------- -With LLM-as-judge evaluators, it is important to carefully review the resulting scores and tune the grader prompt if needed. Often it is helpful to write these as few-shot evaluators, where you provide examples of inputs, outputs, and expected grades as part of the grader prompt. +Any time you pass your dataset’s inputs into your application—whether you’re testing a new prompt, a new model, or a new system configuration—you’re effectively starting an experiment. LangSmith keeps track of these experiments so you can compare differences in outputs side by side. This makes it easier to catch regressions, confirm improvements, and refine your system step by step. -Learn about [how to define an LLM-as-a-judge evaluator](/evaluation/how_to_guides/llm_as_judge). +![Experiment](./static/comparing_multiple_experiments.png) -### Pairwise +Annotation Queues +----------------- -Pairwise evaluators allow you to compare the outputs of two versions of an application. -Think [LMSYS Chatbot Arena](https://chat.lmsys.org/) - this is the same concept, but applied to AI applications more generally, not just models! -This can use either a heuristic ("which response is longer"), an LLM (with a specific pairwise prompt), or human (asking them to manually annotate examples). +Annotation queues power the process of collecting real user feedback. They let you direct runs into a pipeline where human annotators can label, grade, or comment on outputs. You might label every run, or just a sample if your traffic is large. Over time, these labels can form their own dataset for further offline evaluation. Annotation queues are thus a key tool for harnessing human feedback in a consistent, transparent manner. -**When should you use pairwise evaluation?** +To learn more about annotation queues, visit [here](/evaluation/how_to_guides#annotation-queues-and-human-feedback) -Pairwise evaluation is helpful when it is difficult to directly score an LLM output, but easier to compare two outputs. -This can be the case for tasks like summarization - it may be hard to give a summary an absolute score, but easy to choose which of two summaries is more informative. +Offline Evaluation +------------------ -Learn [how run pairwise evaluations](/evaluation/how_to_guides/evaluate_pairwise). - -## Experiment - -Each time we evaluate an application on a dataset, we are conducting an experiment. -An experiment is a single execution of the example inputs in your dataset through your application. -Typically, we will run multiple experiments on a given dataset, testing different configurations of our application (e.g., different prompts or LLMs). -In LangSmith, you can easily view all the experiments associated with your dataset. -Additionally, you can [compare multiple experiments in a comparison view](/evaluation/how_to_guides/compare_experiment_results). - -![Example](./static/comparing_multiple_experiments.png) - -## Annotation queues - -Human feedback is often the most valuable feedback you can gather on your application. -With [annotation queues](/evaluation/how_to_guides/annotation_queues) you can flag runs of your application for annotation. -Human annotators then have a streamlined view to review and provide feedback on the runs in a queue. -Often (some subset of) these annotated runs are then transferred to a [dataset](/evaluation/concepts#datasets) for future evaluations. -While you can always [annotate runs inline](/evaluation/how_to_guides/annotate_traces_inline), annotation queues provide another option to group runs together, specify annotation criteria, and configure permissions. - -Learn more about [annotation queues and human feedback](/evaluation/how_to_guides#annotation-queues-and-human-feedback). - -## Offline evaluation - -Evaluating an application on a dataset is what we call "offline" evaluation. -It is offline because we're evaluating on a pre-compiled set of data. -An online evaluation, on the other hand, is one in which we evaluate a deployed application's outputs on real traffic, in near realtime. -Offline evaluations are used for testing a version(s) of your application pre-deployment. - -You can run offline evaluations client-side using the LangSmith SDK ([Python](https://docs.smith.langchain.com/reference/python) and TypeScript). You can run them server-side via the [Prompt Playground](../../prompt_engineering/concepts#prompt-playground) or by configuring [automations](/observability/how_to_guides/monitoring/rules) to run certain evaluators on every new experiment against a specific dataset. +Offline evaluation focuses on a static dataset rather than live user queries. It’s an excellent practice to verify changes before deployment or measure how your system handles historical use cases. ![Offline](./static/offline.png) ### Benchmarking -Perhaps the most common type of offline evaluation is one in which we curate a dataset of representative inputs, define the key performance metrics, and benchmark multiple versions of our application to find the best one. -Benchmarking can be laborious because for many use cases you have to curate a dataset with gold-standard reference outputs and design good metrics for comparing experimental outputs to them. -For a RAG Q&A bot this might look like a dataset of questions and reference answers, and an LLM-as-judge evaluator that determines if the actual answer is semantically equivalent to the reference answer. -For a ReACT agent this might look like a dataset of user requests and a reference set of all the tool calls the model is supposed to make, and a heuristic evaluator that checks if all of the reference tool calls were made. +Benchmarking compares your system’s outputs to some fixed standard. For question-answering tasks, you might compare the model’s responses against reference answers and compute similarity. Or you could use an LLM-as-judge approach. Typically, large-scale benchmarking is reserved for major system updates, since it requires maintaining extensive curated datasets. -### Unit tests +### Unit Tests -Unit tests are used in software development to verify the correctness of individual system components. -[Unit tests in the context of LLMs are often rule-based assertions](https://hamel.dev/blog/posts/evals/#level-1-unit-tests) on LLM inputs or outputs (e.g., checking that LLM-generated code can be compiled, JSON can be loaded, etc.) that validate basic functionality. +Classic “unit tests” can still be applied to LLMs. You can write logic-based checks looking for empty strings, invalid JSON, or other fundamental errors. These tests can run during your continuous integration (CI) process, catching critical issues anytime you change prompts, models, or other code. -Unit tests are often written with the expectation that they should always pass. -These types of tests are nice to run as part of CI. -Note that when doing so it is useful to set up a cache to minimize LLM calls (because those can quickly rack up!). +### Regression Tests -### Regression tests +Regression tests help ensure that today’s improvements don’t break yesterday’s successes. After a prompt tweak or model update, you can re-run the same dataset and directly compare new results against old ones. LangSmith’s dashboard highlights any degradations in red and improvements in green, making it easy to see how the changes affect overall performance. -Regression tests are used to measure performance across versions of your application over time. -They are used to, at the very least, ensure that a new app version does not regress on examples that your current version correctly handles, and ideally to measure how much better your new version is relative to the current. -Often these are triggered when you are making app updates (e.g. updating models or architectures) that are expected to influence the user experience. - -LangSmith's comparison view has native support for regression testing, allowing you to quickly see examples that have changed relative to the baseline. -Regressions are highlighted red, improvements green. - -![Regression](./static/regression.png) +Illustration: Regression view highlights newly broken examples in red, improvements in green. ### Backtesting -Backtesting is an approach that combines dataset creation (discussed above) with evaluation. If you have a collection of production logs, you can turn them into a dataset. Then, you can re-run those production examples with newer application versions. This allows you to assess performance on past and realistic user inputs. - -This is commonly used to evaluate new model versions. -Anthropic dropped a new model? No problem! Grab the 1000 most recent runs through your application and pass them through the new model. -Then compare those results to what actually happened in production. - -### Pairwise evaluation +Backtesting replays past production runs against your updated system. By comparing new outputs to what you served previously, you gain a real-world perspective on whether the upgrade will solve user pain points or potentially introduce new problems—all without impacting live users. -For some tasks [it is easier](https://www.oreilly.com/radar/what-we-learned-from-a-year-of-building-with-llms-part-i/) for a human or LLM grader to determine if "version A is better than B" than to assign an absolute score to either A or B. -Pairwise evaluations are just this — a scoring of the outputs of two versions against each other as opposed to against some reference output or absolute criteria. -Pairwise evaluations are often useful when using LLM-as-judge evaluators on more general tasks. -For example, if you have a summarizer application, it may be easier for an LLM-as-judge to determine "Which of these two summaries is more clear and concise?" than to give an absolute score like "Give this summary a score of 1-10 in terms of clarity and concision." +### Pairwise Evaluation -Learn [how run pairwise evaluations](/evaluation/how_to_guides/evaluate_pairwise). +Sometimes it’s more natural to decide which output is better rather than relying on absolute scoring. With offline pairwise evaluation, you run both system versions on the same set of inputs and directly compare each example’s outputs. This is commonly used for tasks such as summarization, where multiple outputs may be valid but differ in overall quality. -## Online evaluation +Online Evaluation +----------------- -Evaluating a deployed application's outputs in (roughly) realtime is what we call "online" evaluation. -In this case there is no dataset involved and no possibility of reference outputs — we're running evaluators on real inputs and real outputs as they're produced. -This is be useful for monitoring your application and flagging unintended behavior. -Online evaluation can also work hand-in-hand with offline evaluation: for example, an online evaluator can be used to classify input questions into a set of categories that can be later used to curate a dataset for offline evaluation. - -Online evaluators are generally intended to be run server-side. LangSmith has built-in [LLM-as-judge evaluators](/evaluation/how_to_guides/llm_as_judge) that you can configure, or you can define custom code evaluators that are also run within LangSmith. +Online evaluation measures performance in production, giving you near real-time feedback on potential issues. Instead of waiting for a batch evaluation to conclude, you can detect errors or regressions as soon as they arise. This immediate visibility can be achieved through heuristic checks, LLM-based evaluators, or any custom logic you deploy alongside your live application. ![Online](./static/online.png) -## Application-specific techniques +Application-Specific Techniques +------------------------------- -Below, we will discuss evaluation of a few specific, popular LLM applications. +LangSmith evaluations can be tailored to fit a variety of common LLM application patterns. Below are some popular scenarios and potential evaluation approaches. ### Agents -[LLM-powered autonomous agents](https://lilianweng.github.io/posts/2023-06-23-agent/) combine three components (1) Tool calling, (2) Memory, and (3) Planning. Agents [use tool calling](https://python.langchain.com/v0.1/docs/modules/agents/agent_types/tool_calling/) with planning (e.g., often via prompting) and memory (e.g., often short-term message history) to generate responses. [Tool calling](https://python.langchain.com/v0.1/docs/modules/model_io/chat/function_calling/) allows a model to respond to a given prompt by generating two things: (1) a tool to invoke and (2) the input arguments required. - -![Tool use](./static/tool_use.png) - -Below is a tool-calling agent in [LangGraph](https://langchain-ai.github.io/langgraph/tutorials/introduction/). The `assistant node` is an LLM that determines whether to invoke a tool based upon the input. The `tool condition` sees if a tool was selected by the `assistant node` and, if so, routes to the `tool node`. The `tool node` executes the tool and returns the output as a tool message to the `assistant node`. This loop continues until as long as the `assistant node` selects a tool. If no tool is selected, then the agent directly returns the LLM response. - -![Agent](./static/langgraph_agent.png) +Agents use an LLM to manage decisions, often with access to external tools and memory. Agents break problems into multiple steps, deciding whether to call a tool, how to parse user instructions, and how to proceed based on the results of prior steps. -This sets up three general types of agent evaluations that users are often interested in: +You can assess agents in several ways: -- `Final Response`: Evaluate the agent's final response. -- `Single step`: Evaluate any agent step in isolation (e.g., whether it selects the appropriate tool). -- `Trajectory`: Evaluate whether the agent took the expected path (e.g., of tool calls) to arrive at the final answer. +- **Final Response**. Measure the correctness or helpfulness of the final answer alone, ignoring intermediate steps. +- **Single Step**. Look at each decision in isolation to catch small mistakes earlier in the process. +- **Trajectory**. Examine the agent’s entire chain of actions to see whether it deployed the correct tools or if a suboptimal decision early on led to overall failure. -![Agent-eval](./static/agent_eval.png) +#### Evaluating an Agent’s Final Response -Below we will cover what these are, the components (inputs, outputs, evaluators) needed for each one, and when you should consider this. -Note that you likely will want to do multiple (if not all!) of these types of evaluations - they are not mutually exclusive! +If your main concern is whether the agent’s end answer is correct, you can evaluate it as you would any LLM output. This avoids complexity but may not show where a chain of reasoning went awry. -#### Evaluating an agent's final response +#### Evaluating a Single Step -One way to evaluate an agent is to assess its overall performance on a task. This basically involves treating the agent as a black box and simply evaluating whether or not it gets the job done. +Agents can make multiple decisions in a single run. Evaluating each step separately allows you to spot incremental errors. This approach requires storing detailed run histories for each choice or tool invocation. -The inputs should be the user input and (optionally) a list of tools. In some cases, tool are hardcoded as part of the agent and they don't need to be passed in. In other cases, the agent is more generic, meaning it does not have a fixed set of tools and tools need to be passed in at run time. +#### Evaluating an Agent’s Trajectory -The output should be the agent's final response. +A trajectory-based approach looks at the entire flow, from the initial prompt to the final answer. This might involve comparing the agent’s chain of tool calls to a known “ideal” chain or having an LLM or human reviewer judge the agent’s reasoning. It’s the most thorough method but also the most involved to set up. -The evaluator varies depending on the task you are asking the agent to do. Many agents perform a relatively complex set of steps and the output a final text response. Similar to RAG, LLM-as-judge evaluators are often effective for evaluation in these cases because they can assess whether the agent got a job done directly from the text response. +### Retrieval Augmented Generation (RAG) -However, there are several downsides to this type of evaluation. First, it usually takes a while to run. Second, you are not evaluating anything that happens inside the agent, so it can be hard to debug when failures occur. Third, it can sometimes be hard to define appropriate evaluation metrics. +RAG systems fetch context or documentation from external sources to shape the LLM’s output. These are often used for Q&A applications, enterprise searches, or knowledge-based interactions. -#### Evaluating a single step of an agent - -Agents generally perform multiple actions. While it is useful to evaluate them end-to-end, it can also be useful to evaluate these individual actions. This generally involves evaluating a single step of the agent - the LLM call where it decides what to do. - -The inputs should be the input to a single step. Depending on what you are testing, this could just be the raw user input (e.g., a prompt and / or a set of tools) or it can also include previously completed steps. - -The outputs are just the output of that step, which is usually the LLM response. The LLM response often contains tool calls, indicating what action the agent should take next. - -The evaluator for this is usually some binary score for whether the correct tool call was selected, as well as some heuristic for whether the input to the tool was correct. The reference tool can be simply specified as a string. - -There are several benefits to this type of evaluation. It allows you to evaluate individual actions, which lets you hone in where your application may be failing. They are also relatively fast to run (because they only involve a single LLM call) and evaluation often uses simple heuristic evaluation of the selected tool relative to the reference tool. One downside is that they don't capture the full agent - only one particular step. Another downside is that dataset creation can be challenging, particular if you want to include past history in the agent input. It is pretty easy to generate a dataset for steps early on in an agent's trajectory (e.g., this may only include the input prompt), but it can be difficult to generate a dataset for steps later on in the trajectory (e.g., including numerous prior agent actions and responses). - -#### Evaluating an agent's trajectory - -Evaluating an agent's trajectory involves evaluating all the steps an agent took. - -The inputs are again the inputs to the overall agent (the user input, and optionally a list of tools). - -The outputs are a list of tool calls, which can be formulated as an "exact" trajectory (e.g., an expected sequence of tool calls) or simply a set of tool calls that are expected (in any order). - -The evaluator here is some function over the steps taken. Assessing the "exact" trajectory can use a single binary score that confirms an exact match for each tool name in the sequence. This is simple, but has some flaws. Sometimes there can be multiple correct paths. This evaluation also does not capture the difference between a trajectory being off by a single step versus being completely wrong. - -To address these flaws, evaluation metrics can focused on the number of "incorrect" steps taken, which better accounts for trajectories that are close versus ones that deviate significantly. Evaluation metrics can also focus on whether all of the expected tools are called in any order. - -However, none of these approaches evaluate the input to the tools; they only focus on the tools selected. In order to account for this, another evaluation technique is to pass the full agent's trajectory (along with a reference trajectory) as a set of messages (e.g., all LLM responses and tool calls) an LLM-as-judge. This can evaluate the complete behavior of the agent, but it is the most challenging reference to compile (luckily, using a framework like LangGraph can help with this!). Another downside is that evaluation metrics can be somewhat tricky to come up with. - -### Retrieval augmented generation (RAG) - -Retrieval Augmented Generation (RAG) is a powerful technique that involves retrieving relevant documents based on a user's input and passing them to a language model for processing. RAG enables AI applications to generate more informed and context-aware responses by leveraging external knowledge. - -:::info - -For a comprehensive review of RAG concepts, see our [`RAG From Scratch` series](https://github.com/langchain-ai/rag-from-scratch). - -::: +Comprehensive details on building RAG systems can be found here: +https://github.com/langchain-ai/rag-from-scratch #### Dataset -When evaluating RAG applications, a key consideration is whether you have (or can easily obtain) reference answers for each input question. Reference answers serve as ground truth for assessing the correctness of the generated responses. However, even in the absence of reference answers, various evaluations can still be performed using reference-free RAG evaluation prompts (examples provided below). +For RAG, you typically have queries (and possibly reference answers) in your dataset. With reference answers, offline evaluations can measure how accurately your final output matches the ground truth. Even without reference answers, you can still evaluate by checking whether retrieved documents are relevant and whether the system’s answer is faithful to those documents. #### Evaluator -`LLM-as-judge` is a commonly used evaluator for RAG because it's an effective way to evaluate factual accuracy or consistency between texts. - -![rag-types.png](./static/rag-types.png) - -When evaluating RAG applications, you can have evaluators that require reference outputs and those that don't: - -1. **Require reference output**: Compare the RAG chain's generated answer or retrievals against a reference answer (or retrievals) to assess its correctness. -2. **Don't require reference output**: Perform self-consistency checks using prompts that don't require a reference answer (represented by orange, green, and red in the above figure). - -#### Applying RAG Evaluation - -When applying RAG evaluation, consider the following approaches: - -1. `Offline evaluation`: Use offline evaluation for any prompts that rely on a reference answer. This is most commonly used for RAG answer correctness evaluation, where the reference is a ground truth (correct) answer. - -2. `Online evaluation`: Employ online evaluation for any reference-free prompts. This allows you to assess the RAG application's performance in real-time scenarios. - -3. `Pairwise evaluation`: Utilize pairwise evaluation to compare answers produced by different RAG chains. This evaluation focuses on user-specified criteria (e.g., answer format or style) rather than correctness, which can be evaluated using self-consistency or a ground truth reference. - -#### RAG evaluation summary - -| Evaluator | Detail | Needs reference output | LLM-as-judge? | Pairwise relevant | -| ------------------- | ------------------------------------------------- | ---------------------- | ------------------------------------------------------------------------------------- | ----------------- | -| Document relevance | Are documents relevant to the question? | No | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/rag-document-relevance) | No | -| Answer faithfulness | Is the answer grounded in the documents? | No | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/rag-answer-hallucination) | No | -| Answer helpfulness | Does the answer help address the question? | No | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/rag-answer-helpfulness) | No | -| Answer correctness | Is the answer consistent with a reference answer? | Yes | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/rag-answer-vs-reference) | No | -| Pairwise comparison | How do multiple answer versions compare? | No | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/pairwise-evaluation-rag) | Yes | +RAG evaluators commonly focus on factual correctness and faithfulness to the retrieved information. You can carry out these checks offline (with reference answers), online (in near real-time for live queries), or in pairwise comparisons (to compare different ranking or retrieval methods). ### Summarization -Summarization is one specific type of free-form writing. The evaluation aim is typically to examine the writing (summary) relative to a set of criteria. - -`Developer curated examples` of texts to summarize are commonly used for evaluation (see a dataset example [here](https://smith.langchain.com/public/659b07af-1cab-4e18-b21a-91a69a4c3990/d)). However, `user logs` from a production (summarization) app can be used for online evaluation with any of the `Reference-free` evaluation prompts below. - -`LLM-as-judge` is typically used for evaluation of summarization (as well as other types of writing) using `Reference-free` prompts that follow provided criteria to grade a summary. It is less common to provide a particular `Reference` summary, because summarization is a creative task and there are many possible correct answers. - -`Online` or `Offline` evaluation are feasible because of the `Reference-free` prompt used. `Pairwise` evaluation is also a powerful way to perform comparisons between different summarization chains (e.g., different summarization prompts or LLMs): - -| Use Case | Detail | Needs reference output | LLM-as-judge? | Pairwise relevant | -| ---------------- | -------------------------------------------------------------------------- | ---------------------- | -------------------------------------------------------------------------------------------- | ----------------- | -| Factual accuracy | Is the summary accurate relative to the source documents? | No | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/summary-accurancy-evaluator) | Yes | -| Faithfulness | Is the summary grounded in the source documents (e.g., no hallucinations)? | No | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/summary-hallucination-evaluator) | Yes | -| Helpfulness | Is summary helpful relative to user need? | No | Yes - [prompt](https://smith.langchain.com/hub/langchain-ai/summary-helpfulness-evaluator) | Yes | +Summarization tasks are often subjective, making it challenging to define a single “correct” output. In this context, LLM-as-judge strategies are particularly useful. By asking a language model to grade clarity, accuracy, or coverage, you can track your summarizer’s performance. Alternatively, offline pairwise comparisons can help you see which summary outperforms the other—especially if you’re testing new prompt styles or models. ### Classification / Tagging -Classification / Tagging applies a label to a given input (e.g., for toxicity detection, sentiment analysis, etc). Classification / Tagging evaluation typically employs the following components, which we will review in detail below: - -A central consideration for Classification / Tagging evaluation is whether you have a dataset with `reference` labels or not. If not, users frequently want to define an evaluator that uses criteria to apply label (e.g., toxicity, etc) to an input (e.g., text, user-question, etc). However, if ground truth class labels are provided, then the evaluation objective is focused on scoring a Classification / Tagging chain relative to the ground truth class label (e.g., using metrics such as precision, recall, etc). - -If ground truth reference labels are provided, then it's common to simply define a [custom heuristic evaluator](./how_to_guides/custom_evaluator) to compare ground truth labels to the chain output. However, it is increasingly common given the emergence of LLMs simply use `LLM-as-judge` to perform the Classification / Tagging of an input based upon specified criteria (without a ground truth reference). - -`Online` or `Offline` evaluation is feasible when using `LLM-as-judge` with the `Reference-free` prompt used. In particular, this is well suited to `Online` evaluation when a user wants to tag / classify application input (e.g., for toxicity, etc). +Classification tasks apply labels or tags to inputs. If you have reference labels, you can compute metrics like accuracy, precision, or recall. If not, you can still apply LLM-as-judge techniques, instructing the model to validate whether a predicted label matches labeling guidelines. Pairwise evaluation is also an option if you need to compare two classification systems. -| Use Case | Detail | Needs reference output | LLM-as-judge? | Pairwise relevant | -| --------- | ------------------- | ---------------------- | ------------- | ----------------- | -| Accuracy | Standard definition | Yes | No | No | -| Precision | Standard definition | Yes | No | No | -| Recall | Standard definition | Yes | No | No | +In all these application patterns, LangSmith’s offline and online tools—and the combination of heuristics, LLM-based evaluations, human feedback, and pairwise comparisons—can help maintain and improve performance as your system evolves. \ No newline at end of file