Fix dead neptune.ai llm-observability citation links

content/faq/all/ai-research-assistant-monitoring.mdx

@@ -87,7 +87,7 @@ Logging raw data is not enough – you also need to aggregate and visualize it.
 - **Latency and Throughput:** Plot how long the agent takes to respond and how that changes with load. If response time is creeping up as more users join, you might need to optimize or scale your infrastructure.
 - **Cost Breakdown:** If your agent uses a paid API (like OpenAI) or other billable resources, monitor cost per query and which features incur the most cost. For instance, you may discover that one type of analysis the agent does (like an exhaustive data summary) costs 5× more than a simpler Q&A – information that could guide when to enable or disable certain expensive operations.
 
-Langfuse provides built-in dashboards for these metrics, so you can see things like average response time, tokens used, and even custom scores all in one place ([LLM Observability: Fundamentals, Practices, and Tools](https://neptune.ai/blog/llm-observability#:~:text=,It%20visualizes%20costs)) ([LLM Observability: Fundamentals, Practices, and Tools](https://neptune.ai/blog/llm-observability#:~:text=match%20at%20L723%20,of%20input%20and%20output%20tokens)). This consolidated view helps **evaluate your AI agent at a high level** and communicate its performance to your team. (It’s much easier to show a graph of “accuracy over time” in a meeting than to explain a bunch of log lines!)
+Langfuse provides built-in dashboards for these metrics, so you can see things like average response time, tokens used, and even custom scores all in one place ([LLM Observability: Fundamentals, Practices, and Tools](https://web.archive.org/web/20251005112019/https://neptune.ai/blog/llm-observability#:~:text=,It%20visualizes%20costs)) ([LLM Observability: Fundamentals, Practices, and Tools](https://web.archive.org/web/20251005112019/https://neptune.ai/blog/llm-observability#:~:text=match%20at%20L723%20,of%20input%20and%20output%20tokens)). This consolidated view helps **evaluate your AI agent at a high level** and communicate its performance to your team. (It’s much easier to show a graph of “accuracy over time” in a meeting than to explain a bunch of log lines!)
 
 ### 4. Continuously Test and **Evaluate** the AI Agent
 
@@ -99,7 +99,7 @@ Important practices include:
 - **User Feedback Loop:** Once deployed, collect feedback from actual users. If a scientist using the assistant flags an answer as incorrect, that should be logged and fed back into an evaluation set. Over time, you build a dataset of question→expected answer pairs (or ratings) that you can use to measure the agent’s quality. This is essentially creating an evaluation benchmark that grows with production data.
 - **Regular Benchmarking:** As you update the agent (maybe you fine-tune the model or add a new data source), re-run your evaluation set and compare results. Did the accuracy improve in summarizing biology papers but worsen for economics papers? Continuous evaluation helps catch regressions. It’s similar to how software tests ensure new code doesn’t break old functionality – here we ensure a new model or prompt doesn’t break the agent’s reliability.
 
-Langfuse makes this easier with features for automated evaluations and dataset management. You can log expected outputs and compare them with the agent's actual outputs, using either statistical metrics or even other AI models to score the answers. In fact, Langfuse supports a **wide range of evaluation approaches**, including automated model-based scoring and direct human annotations of the AI’s outputs ([LLM Observability: Fundamentals, Practices, and Tools](https://neptune.ai/blog/llm-observability#:~:text=,directly%20in%20the%20platform%E2%80%99s%20UI)). This means your team can rate an answer as "good" or "bad" right in the monitoring UI, or use an AI evaluator to judge correctness – and all that data gets tracked. Over time, you get a quantitative view of the agent’s performance (for example, "our agent’s answer quality is 8/10 on average this month, up from 7/10 last month after we updated the knowledge base"). Such **AI agent evaluation** is key to confidently scaling usage – you know its limits and strengths through hard data.
+Langfuse makes this easier with features for automated evaluations and dataset management. You can log expected outputs and compare them with the agent's actual outputs, using either statistical metrics or even other AI models to score the answers. In fact, Langfuse supports a **wide range of evaluation approaches**, including automated model-based scoring and direct human annotations of the AI’s outputs ([LLM Observability: Fundamentals, Practices, and Tools](https://web.archive.org/web/20251005112019/https://neptune.ai/blog/llm-observability#:~:text=,directly%20in%20the%20platform%E2%80%99s%20UI)). This means your team can rate an answer as "good" or "bad" right in the monitoring UI, or use an AI evaluator to judge correctness – and all that data gets tracked. Over time, you get a quantitative view of the agent’s performance (for example, "our agent’s answer quality is 8/10 on average this month, up from 7/10 last month after we updated the knowledge base"). Such **AI agent evaluation** is key to confidently scaling usage – you know its limits and strengths through hard data.
 
 ## Langfuse vs. Other Observability Solutions
 
@@ -116,10 +116,10 @@ There are a few ways one might monitor an AI system – from DIY logging setups
 
 **Langfuse’s approach:** Langfuse aims to provide a **unified platform** covering tracing, analytics, and evaluation specifically for LLM applications. A few key features that set it apart include:
 
-- **Prompt Management and Versioning:** Langfuse includes a prompt management system (a "Prompt CMS") to version prompts and track changes in prompt performance ([LLM Observability: Fundamentals, Practices, and Tools](https://neptune.ai/blog/llm-observability#:~:text=,explicit%20and%20implicit%20user%20feedback)). This is crucial when your research agent’s effectiveness can depend heavily on how you phrase its instructions. You can experiment with prompt tweaks and see differences in outcomes, all while keeping a history.
-- **Integrated Tracing:** Langfuse offers out-of-the-box integrations with popular AI frameworks (LangChain, LlamaIndex, etc.) to capture traces effortlessly ([LLM Observability: Fundamentals, Practices, and Tools](https://neptune.ai/blog/llm-observability#:~:text=,recorded%20traces%20via%20an%20API)). It visualizes the agent’s chain-of-thought in an intuitive way. You don’t have to manually instrument every step – use the Langfuse SDK, and it will hook into the model calls, tool usage, and decisions automatically.
-- **Usage & Cost Monitoring:** Langfuse provides real-time monitoring of usage metrics and costs. It can retrieve API cost reports or infer them based on token counts, giving you a live tally of how many dollars your agent’s research spree is costing ([LLM Observability: Fundamentals, Practices, and Tools](https://neptune.ai/blog/llm-observability#:~:text=,of%20input%20and%20output%20tokens)). This is especially handy for businesses to **manage the budget** on AI usage.
-- **Evaluation Suite:** As mentioned, Langfuse has a built-in evaluation workflow. You can create evaluation datasets and track scores over time, perform A/B tests between different agent versions, and even collect user feedback directly through the platform ([LLM Observability: Fundamentals, Practices, and Tools](https://neptune.ai/blog/llm-observability#:~:text=,directly%20in%20the%20platform%E2%80%99s%20UI)). Instead of juggling separate tools for monitoring vs. evaluation, Langfuse combines them, which streamlines **AI agent evaluation** efforts.
+- **Prompt Management and Versioning:** Langfuse includes a prompt management system (a "Prompt CMS") to version prompts and track changes in prompt performance ([LLM Observability: Fundamentals, Practices, and Tools](https://web.archive.org/web/20251005112019/https://neptune.ai/blog/llm-observability#:~:text=,explicit%20and%20implicit%20user%20feedback)). This is crucial when your research agent’s effectiveness can depend heavily on how you phrase its instructions. You can experiment with prompt tweaks and see differences in outcomes, all while keeping a history.
+- **Integrated Tracing:** Langfuse offers out-of-the-box integrations with popular AI frameworks (LangChain, LlamaIndex, etc.) to capture traces effortlessly ([LLM Observability: Fundamentals, Practices, and Tools](https://web.archive.org/web/20251005112019/https://neptune.ai/blog/llm-observability#:~:text=,recorded%20traces%20via%20an%20API)). It visualizes the agent’s chain-of-thought in an intuitive way. You don’t have to manually instrument every step – use the Langfuse SDK, and it will hook into the model calls, tool usage, and decisions automatically.
+- **Usage & Cost Monitoring:** Langfuse provides real-time monitoring of usage metrics and costs. It can retrieve API cost reports or infer them based on token counts, giving you a live tally of how many dollars your agent’s research spree is costing ([LLM Observability: Fundamentals, Practices, and Tools](https://web.archive.org/web/20251005112019/https://neptune.ai/blog/llm-observability#:~:text=,of%20input%20and%20output%20tokens)). This is especially handy for businesses to **manage the budget** on AI usage.
+- **Evaluation Suite:** As mentioned, Langfuse has a built-in evaluation workflow. You can create evaluation datasets and track scores over time, perform A/B tests between different agent versions, and even collect user feedback directly through the platform ([LLM Observability: Fundamentals, Practices, and Tools](https://web.archive.org/web/20251005112019/https://neptune.ai/blog/llm-observability#:~:text=,directly%20in%20the%20platform%E2%80%99s%20UI)). Instead of juggling separate tools for monitoring vs. evaluation, Langfuse combines them, which streamlines **AI agent evaluation** efforts.
 
 In summary, alternatives like Helicone or Phoenix might excel in one dimension (simple setup or advanced ML analytics, respectively), but Langfuse’s strength is in providing a balanced, all-in-one observability solution for AI agents. And since Langfuse is open-source, you have flexibility to self-host for privacy or extend it to your needs, something closed solutions might not offer. As one comparison noted, _teams seeking an open-source alternative for LLM observability_ often choose Langfuse for its transparency and powerful feature set ([Compare: The Best LangSmith Alternatives & Competitors - Helicone](https://www.helicone.ai/blog/best-langsmith-alternatives#:~:text=Langfuse%20can%20be%20more%20cost,AI%20excels%20in%20those)).