Building an LLM evaluation harness
Automated LLM testing. This post digs into the topic, drawing on recent discussion from Hacker News, Reddit, and the wider community.
The conversation that prompted this post is linked in the sources. I have tried to capture the substance of the discussion rather than just summarize it, because the comments are usually where the real insight lives.
What happened
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Prompt engineering is real engineering. The difference between a mediocre and an excellent prompt is often the difference between a demo and a product, and it deserves the same rigor as code.
Prompt engineering is real engineering. The difference between a mediocre and an excellent prompt is often the difference between a demo and a product, and it deserves the same rigor as code.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Tool use turns a chat model into an agent. The model decides which tool to call, with what arguments, and interprets the result. Getting this loop right is where most of the engineering effort goes.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Structured output via JSON schema constraints makes LLMs usable as components in larger systems. Unstructured text is a liability in a pipeline.
Caching is the single biggest cost lever for LLM applications. Prompt caching and response caching can cut bills by an order of magnitude for workloads with repeated prefixes.
The hardest part of building with LLMs is not the model call. It is the evaluation. Without a solid eval suite, you cannot tell whether a change helped or hurt, and you end up shipping on vibes.
Why it matters
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Large language models are best understood as very good pattern matchers with a large but finite context. They are not databases, not search engines, and not oracles, and treating them as any of those leads to predictable failures.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Large language models are best understood as very good pattern matchers with a large but finite context. They are not databases, not search engines, and not oracles, and treating them as any of those leads to predictable failures.
Fine-tuning is overused. Most teams should exhaust prompting and RAG before considering a fine-tune, because fine-tuning is expensive to evaluate and hard to iterate on.
Observability for LLM apps means logging prompts, completions, tool calls, and costs. Without it you are flying blind.
The hardest part of building with LLMs is not the model call. It is the evaluation. Without a solid eval suite, you cannot tell whether a change helped or hurt, and you end up shipping on vibes.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Caching is the single biggest cost lever for LLM applications. Prompt caching and response caching can cut bills by an order of magnitude for workloads with repeated prefixes.
Evaluation is a product, not a test. A good eval suite encodes what your users actually care about and becomes the spec your model is measured against.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Evaluation is a product, not a test. A good eval suite encodes what your users actually care about and becomes the spec your model is measured against.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Practical takeaways
- Latency is a feature.
- Retrieval quality dominates RAG quality.
- Guardrails are not optional in production.
- Context windows keep growing, but attention is not free.
- The hardest part of building with LLMs is not the model call.
- Guardrails are not optional in production.
- Large language models are best understood as very good pattern matchers with a large but finite context.
- Caching is the single biggest cost lever for LLM applications.
- Prompt engineering is real engineering.
- Structured output via JSON schema constraints makes LLMs usable as components in larger systems.
Evaluation is a product, not a test. A good eval suite encodes what your users actually care about and becomes the spec your model is measured against.
Fine-tuning is overused. Most teams should exhaust prompting and RAG before considering a fine-tune, because fine-tuning is expensive to evaluate and hard to iterate on.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Observability for LLM apps means logging prompts, completions, tool calls, and costs. Without it you are flying blind.
Caching is the single biggest cost lever for LLM applications. Prompt caching and response caching can cut bills by an order of magnitude for workloads with repeated prefixes.
The hardest part of building with LLMs is not the model call. It is the evaluation. Without a solid eval suite, you cannot tell whether a change helped or hurt, and you end up shipping on vibes.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
The hardest part of building with LLMs is not the model call. It is the evaluation. Without a solid eval suite, you cannot tell whether a change helped or hurt, and you end up shipping on vibes.
Tool use turns a chat model into an agent. The model decides which tool to call, with what arguments, and interprets the result. Getting this loop right is where most of the engineering effort goes.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Evaluation is a product, not a test. A good eval suite encodes what your users actually care about and becomes the spec your model is measured against.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
The hardest part of building with LLMs is not the model call. It is the evaluation. Without a solid eval suite, you cannot tell whether a change helped or hurt, and you end up shipping on vibes.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Context windows keep growing, but attention is not free. Stuffing more tokens into the prompt dilutes the signal and increases cost and latency. Retrieval and summarization usually beat brute force.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Sources
This post is informed by discussion on github.com and related threads on Hacker News and Reddit. Primary source: https://github.com/confident-ai/deepeval.
Evaluation is a product, not a test. A good eval suite encodes what your users actually care about and becomes the spec your model is measured against.
Prompt engineering is real engineering. The difference between a mediocre and an excellent prompt is often the difference between a demo and a product, and it deserves the same rigor as code.
Context windows keep growing, but attention is not free. Stuffing more tokens into the prompt dilutes the signal and increases cost and latency. Retrieval and summarization usually beat brute force.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Evaluation is a product, not a test. A good eval suite encodes what your users actually care about and becomes the spec your model is measured against.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Evaluation is a product, not a test. A good eval suite encodes what your users actually care about and becomes the spec your model is measured against.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Fine-tuning is overused. Most teams should exhaust prompting and RAG before considering a fine-tune, because fine-tuning is expensive to evaluate and hard to iterate on.
Tool use turns a chat model into an agent. The model decides which tool to call, with what arguments, and interprets the result. Getting this loop right is where most of the engineering effort goes.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Caching is the single biggest cost lever for LLM applications. Prompt caching and response caching can cut bills by an order of magnitude for workloads with repeated prefixes.
Observability for LLM apps means logging prompts, completions, tool calls, and costs. Without it you are flying blind.
Tool use turns a chat model into an agent. The model decides which tool to call, with what arguments, and interprets the result. Getting this loop right is where most of the engineering effort goes.
Caching is the single biggest cost lever for LLM applications. Prompt caching and response caching can cut bills by an order of magnitude for workloads with repeated prefixes.
Large language models are best understood as very good pattern matchers with a large but finite context. They are not databases, not search engines, and not oracles, and treating them as any of those leads to predictable failures.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Caching is the single biggest cost lever for LLM applications. Prompt caching and response caching can cut bills by an order of magnitude for workloads with repeated prefixes.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Context windows keep growing, but attention is not free. Stuffing more tokens into the prompt dilutes the signal and increases cost and latency. Retrieval and summarization usually beat brute force.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Tool use turns a chat model into an agent. The model decides which tool to call, with what arguments, and interprets the result. Getting this loop right is where most of the engineering effort goes.
Observability for LLM apps means logging prompts, completions, tool calls, and costs. Without it you are flying blind.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Prompt engineering is real engineering. The difference between a mediocre and an excellent prompt is often the difference between a demo and a product, and it deserves the same rigor as code.
Structured output via JSON schema constraints makes LLMs usable as components in larger systems. Unstructured text is a liability in a pipeline.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Context windows keep growing, but attention is not free. Stuffing more tokens into the prompt dilutes the signal and increases cost and latency. Retrieval and summarization usually beat brute force.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Structured output via JSON schema constraints makes LLMs usable as components in larger systems. Unstructured text is a liability in a pipeline.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Tool use turns a chat model into an agent. The model decides which tool to call, with what arguments, and interprets the result. Getting this loop right is where most of the engineering effort goes.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Model routing sends simple queries to cheap models and hard ones to expensive models. It is the easiest cost optimization most teams have not done.
Caching is the single biggest cost lever for LLM applications. Prompt caching and response caching can cut bills by an order of magnitude for workloads with repeated prefixes.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Guardrails are not optional in production. Models will eventually produce something you did not want, and the question is whether your system catches it before the user does.
Large language models are best understood as very good pattern matchers with a large but finite context. They are not databases, not search engines, and not oracles, and treating them as any of those leads to predictable failures.
The hardest part of building with LLMs is not the model call. It is the evaluation. Without a solid eval suite, you cannot tell whether a change helped or hurt, and you end up shipping on vibes.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
Tool use turns a chat model into an agent. The model decides which tool to call, with what arguments, and interprets the result. Getting this loop right is where most of the engineering effort goes.
Large language models are best understood as very good pattern matchers with a large but finite context. They are not databases, not search engines, and not oracles, and treating them as any of those leads to predictable failures.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
Context windows keep growing, but attention is not free. Stuffing more tokens into the prompt dilutes the signal and increases cost and latency. Retrieval and summarization usually beat brute force.
Retrieval quality dominates RAG quality. A perfect model with bad retrieval gives bad answers, and a mediocre model with great retrieval gives acceptable ones.
Latency is a feature. Users abandon slow AI features faster than slow traditional ones because the expectation of instant response is higher.
Prompt engineering is real engineering. The difference between a mediocre and an excellent prompt is often the difference between a demo and a product, and it deserves the same rigor as code.
Multi-agent systems are promising but add coordination complexity. Start with a single agent and a good tool set before splitting into specialists.
Structured output via JSON schema constraints makes LLMs usable as components in larger systems. Unstructured text is a liability in a pipeline.
Structured output via JSON schema constraints makes LLMs usable as components in larger systems. Unstructured text is a liability in a pipeline.
Fine-tuning is overused. Most teams should exhaust prompting and RAG before considering a fine-tune, because fine-tuning is expensive to evaluate and hard to iterate on.