Scores & Evaluation
Evaluation is a critical aspect of developing and deploying LLM applications. Usually teams use a multitude of different evaluation methods depending on the use case and the stage of the development process. Langfuse provides a flexible scoring system to capture these evaluations and make them actionable.
Common evaluation methods
Langfuse supports all forms of evaluation methods due to its open architecture and API, learn more on the score object in the next section.
Depending on your needs in the development process, you can use one or multiple of the following evaluation methods. Each method will vary in: frequency, cost, and quality. usually only a combination of multiple methods will provide a comprehensive evaluation of the LLM application.
| Source | examples |
|---|---|
| Manual Annotation (in UI) | Establish an evaluation baseline by manually annotating data in the UI |
| User feedback | Explicit (e.g., thumbs up/down, 1-5 star rating) or implicit (e.g., time spent on a page, click-through rate, accepting/rejecting a model-generated output, human-in-the-loop) |
| Model-based evaluation | Langfuse-managed evals or external library: OpenAI Evals, Whylabs Langkit, Langchain Evaluators (cookbook), RAGAS for RAG pipelines (cookbook), custom model outputs |
| Custom via SDKs/API | Run-time quality checks (e.g. valid structured output format), custom link to annotation tools, external evalutation pipelines that sync with Langfuse. |
Score object in Langfuse
Scores serve as an object to store evaluation metrics in Langfuse. They always are related to a trace and can be attached to specific observations within a trace. Optionally, scores can be related to a score config to ensure they comply with a specific schema.
| Attribute | Type | Description |
|---|---|---|
name | string | Name of the score, e.g. user_feedback, hallucination_eval |
value | number | Optional: Numeric value of the score. Always defined for numeric and boolean scores. Optional for categorical scores. |
stringValue | string | Optional: String equivalent of the score's numeric value for boolean and categorical data types. Automatically set for categorical scores based on the config if the configId is provided. |
traceId | string | Id of the trace the score relates to |
observationId | string | Optional: Observation (e.g. LLM call) the score relates to |
comment | string | Optional: Evaluation comment, commonly used for user feedback, eval output or internal notes |
id | string | Unique identifier of the score. Auto-generated by SDKs. Optionally can also be used as an indempotency key to update scores. |
source | string | Automatically set based on the souce of the score. Can be either API, EVAL, or ANNOTATION |
dataType | string | Automatically set based on the config data type when the configId is provided. Otherwise can be defined manually as NUMERIC, CATEGORICAL or BOOLEAN |
configId | string | Optional: Score config id to ensure that the score follows a specific schema. Can be defined in the Langfuse UI or via API. When provided the score's dataType is automatically set based on the config |
Using scores across Langfuse
Scores can be used in multiple ways across Langfuse:
- Displayed on trace to provide a quick overview
- Segment all execution traces by scores to e.g. find all traces with a low quality score
- Analytics: Detailed score reporting with drill downs into use cases and user segments
Configure score schema
If you'd like to ensure that your scores follow a specific schema, you can define a score config in the Langfuse UI or via our API. A score config includes the score name, data type, and constraints on score value range such as min and max values for numerical data types and custom categories for categorical data types. Configs are immutable but can be archived (and restored anytime). Using score configs allows you to standardize your scoring schema across your team and ensure that scores are consistent and comparable for future analysis.
Frequently used scores
Scores in Langfuse are adaptable (it is just a name) and designed to cater to the unique requirements of specific LLM applications. They typically serve to measure the following aspects:
- Quality
- Factual accuracy
- Completeness of the information provided
- Verification against hallucinations
- Style
- Sentiment portrayed
- Tonality of the content
- Potential toxicity
- Security
- Similarity to prevalent prompt injections
- Instances of model refusals (e.g., as a language model, ...)
This flexible scoring system allows for a comprehensive evaluation of various elements integral to the function and performance of the LLM application.