Here is quick rundown of what we've built in April and May ⬇️

We've heard you, Analytics really needed a revamp. We're releasing for everyone (including free users) the next generation of Analytics.

Find new charts and insights, filter data and select any date range you want. It also introduces an hourly breakdown for more granular insights.

You can now edit the assistant's tools calling content in your prompts, allowing to build Chain-of-thought prompts using tools. This enables many use-cases.

But also:

• 💵 New Pricing: More affordable pricing structure at $20 per seat per month for the cloud version. Also more generous in features.
• 🛣️ Public Roadmap: Follow upcoming improvements on our public roadmap.
• 🗒️ Prompt Notepad: Added a free text field on prompts to provide context and explanations for your prompt designs.
• 🏗️ Direct Fine-tune Exports: Export data directly to a OpenAI-compatible JSONL file, allowing you to fine-tune GPT models in no time.
• ⛩️ Templates: View which template was used for an LLM call as well as filter LLM logs by template. You will now also see in the Analytics section your most used templates.
• 👥 Delete Users: Delete users and associated data in 1-click from their page to easily comply with privacy requests.
• 🤖 GPT-4o: We've added full support for GPT-4o the day it was released.
• 🧬 Tool Calls generator: Easily generate OpenAI tool calling schemas with our public generator.
• 👾 Self-hosting Enhancements: We've refined the Helm charts to be even easier to setup and allowed users to use Azure OpenAI.
• 🪲 Bug Squashing: Too numerous to mention here but our engineering team has been extremely busy fixing bugs all across the board. Stability is a big priority for the near-term for us. If you had an issue with something not working in the past, there are good chances it's resolved.

• ⛑️ Evaluations: We'll be releasing lots of improvements to our Evaluations (and Real-time Evaluations) for better performance, stability and flexibility when working with custom evaluation logics. For example, you'll be able to push the results of your own evaluators to Lunary.
• 🤖 Prompts A/B testing: We'll be pushing tools to easily A/B test prompt versions and determine which version is more efficient.
• 📦 Integrations: We've been slow to add new integrations as they take our team a lot of time to maintain and keep up-to-date, but we'll be ramping up the development of new integrations, starting with LlamaIndex and MotleyCrew.

This month, our focus was on enhancing existing features, particularly evaluations and enterprise-specific functionalities.

• Enhanced search and sorting capabilities.

• Introduced Mistral Large, Medium, and all Claude models.

• Fix issues with Radars not running properly ML models.
• Improved PII and Profanity detection across all languages.

• You can now set and filter by the metadata field within the dashboard for better data tracking.

• Reintroduced the feedback filter.
• Significantly improved the Users filter, including a new search feature for managing large user bases.

• Temporarily disabled last month for security reasons, Private API keys are now re-enabled. Access them in your Settings page. Documentation improvements are forthcoming.

• Enhanced results table for clearer relationships between results.
• Transitioned from models to "Providers" to allow experimentation with parameters like temperature.
• Implemented a queue system for enhanced performance and to address Anthropic rate limiting.
• Added support for running evaluations on large datasets with multiple models.
• Introduced live progress feedback for lengthy evaluations.
• Improved error handling for failed evaluations.

• Deployment made simpler with new Helm Charts for Kubernetes and monolithic Docker images.
• Updated our Role Based Access Control system to allow more granular permission settings.

• JS SDK: Resolved issues with OpenAI streaming types.
• Addressed cascading feedback issues from thread messages.
• Fixed display issues with parameters such as maxTokens.

After the busy last months that resulted in almost an overall make-over of the platform, in February, we've focused on stability improvements across the board. We've also made a number of improvements to the dashboard UI, the evaluations, and the templates.

We've improved the UI of the traces to be more readable and work better with smaller screens.

All Unlimited and Enterprise users now have access to the Evaluation SDK.

Create CI pipelines for your agents, benchmark your RAG pipelines, and more. The SDK is now generally available and ready for production use.

Here is an example of how to use the SDK:

testing_dataset = lunary.get_dataset("test-cases")

for item in testing_dataset:

  # Run your agent on each input of the dataset
  res = support_agent(item.input)

  # Evaluate the output
  passed, details = lunary.evaluate(
    checklist='ci-checks',
    input=item.input,
    output=res,
    ideal_output=item.ideal_output
  )

  if passed:
      print(f"Test passed!", 'green')
  else:
      print(f"Test failed!")

You can now use text-only datasets (instead of chat-messages datasets only before). This makes it easier to create test cases for your custom agents.

For those that work with LangChain, we've added new methods to the SDK that can pull your templates directly as LangChain's ChatPromptTemplate and PromptTemplate classes.

This makes it much easier to work with chains and LangChain in general.

Example:

template = lunary.get_langchain_template("my-template")

messages = lc_template.format_messages(question="What is the capital of France?")

Make sure to update your SDK to the latest version to use these new methods.

Come test your prompts with the new state-of-the-art model. We've found it's better and more concise than GPT-4 in many cases.

We've added tracking of Mistral and Claude 3 prices in the dashboard.

We've removed the dependency on PgGroonga to make it easier to self-host the platform with hosted Postgres services like AWS RDS.

This means Lunary is now also compatible with Postgres 16.

We've increased the number of projects that free users can have from 1 to 3. This should help you better organize your work and keep your projects separate.

In the Python SDK, we've pushed numerous improvements to the parsing of LangChain traces. The constant updates to the LangChain format have made it difficult to keep up with the changes, but we're made significant progress in this area.

We've fixed a number of issues across the platform where:

• Evaluations would not start correctly from the dashboard
• Editing datasets would stutter and not work correctly
• Deleting a template would not work correctly
• Switching between text and chat templates would break the playground
• Invalid tools would prevent the playground from running
• In JS SDK await lunary.flush not working
• JS SDK crashing in old Node environments because of crypto module

January has been our most productive month to date.

We've successfully migrated our entire platform to a new, more efficient architecture. Our new Radars and Evaluations features are now publicly available to all users on Unlimited or self-hosted plans, and we've pushed a lot of usage improvements to the app.

We've revamped our entire source code and eliminated the need for Vercel and Supabase, leading to several improvements:

• Enhanced scalability of the app
• Access to all dashboard data via the API
• 10x increase in our speed of rolling out new features
• Less security concerns and error-surface associated with 3rd-party dependencies
• Simplified self-hosting setup
• Quicker dashboard performance

This is what it looks like in our GitHub graph:

Radars have moved out of private access.

Radars are AI-powered alerts that monitor your runs for specific conditions, such as personal data, profanity, or negative sentiment.

Internally, we've deployed efficient, lightweight AI models that scan runs without relying on external API queries, perfect for self-hosted setups.

We've reimagined LLM evaluations from scratch, opting for no-code approach.

You can design and execute evaluations directly from the dashboard without needing to be a Python expert or have prior evaluation experience.

This is made possible through intuitive blocks that assess various aspects like:

• How closely your model's outputs match an ideal output
• The presence of hallucination in responses
• Usage of costs and tokens

You can create evaluations with 20+ models, including open-source options like Mistral, directly from the dashboard.

Some evaluation scenarios still require code---say for testing your custom agents or integrating into a CI pipeline---we will soon release the Evaluation SDK. This allows for the execution of dashboard-created evals your codebase, enabling contributions from both technical and non-technical team members.

We're continuously refining these features and are open to providing private access to the Evaluations SDK to those interested.

We've pushed a lot of improvements to the app, such as:

• Merged "LLM Calls", "Chats" and "Traces" into a single "Logs" page, improving filters and search functionality across all data types.
• New filtering engine with a lot of additional filters (with more on the way).
• Add tool calls to templates and preview them in your runs.
• Button to duplicate templates
• Improved filter and dashboard speed.
• Simpler navigation around the app

Alongside these updates, we've made numerous fixes and are focused on extending these advancements to more affordable plans as we further optimize and reduce costs.

We're eagerly await your feedback :)

Also, if anyone in San Francisco wants to meet up with the founders, we're here for the month. Hit us up!

Our new Prompt Templates feature is ready in alpha today.

Collaborate with non-technical team members, version prompts and decouple prompts from your source-code.

It's available today to all users, regardless of your plan and the docs can be found here.

JS integration is ready and we'll release the Python integration later this week.

Any feedback on this is greatly appreciated - we will iterate on this in the coming weeks with your feedback.

We heard you and simplified a lot the API to track messages.

You can also now track users' conversations directly from your Python or JS backend, with a much simpler API.

Check out the new Chat tracking docs here.

We now support OpenAI's vision models and you can see the pictures used in your dashboard.

Find Gemini Pro and Mixtral in the prompt playground.

We've turbocharged our dashboard and Postgres database for quicker data rendering. Search, filters and navigation should be much quicker.

It's still a work in progress, but you should already feel the difference with heavy data loads.

There is also a lot more in terms of bug fixed and UI improvements to the dashboard.

Enjoy the holidays!

Big news from our corner! We're shifting gears and our startup's name is changing from LLMonitor to Lunary.ai.

First off - saying 'LLMonitor' wasn't the smoothest. We heard you trying to pronounce it (and saw some of those puzzled looks). So, we're making it easier. Lunary rolls off the tongue way better, doesn't it?

Also, a bit of a hiccup with Google. Our early SEO experiments, let's just say, were a tad too experimental and abrupt. Google wasn't thrilled, and we got a penalty on our main llmonitor.com domain making us impossible to be found on Google. Lesson learned.

But hey, every cloud has a silver lining. This change isn't just about a new name. It's a fresh start, a clearer identity. Lunary.ai reflects our work better - to build the best AI developer platform.

Stay tuned for what's next. We're just getting started.