What's new in Spanloom
Regular updates as we build in public during early access. Every change is shipped to production, tested by real teams, and documented here.
RAG retrieval audit view
Per-request chunk inspection with relevance scores, context utilization rates, and dead-chunk detection. Click any trace row to see exactly which documents your retrieval pipeline fetched and which ones the model actually used.
Prompt diff view
When a prompt version drift alert fires, the new diff view shows a side-by-side comparison between your declared baseline and the live version. Character-level diffs highlight exactly what changed.
Webhook alerts
Configure alert endpoints from the workspace settings. Spanloom sends a signed JSON payload to your webhook URL when latency, error rate, token spend, or prompt hash thresholds are exceeded.
Token count accuracy fix for Claude models
Anthropic's tokenizer for Claude 3.x models counts multi-byte characters differently than GPT-4. Token counts in the cost explorer were off by 3-8% for Claude traces. This is now corrected using the official Anthropic tokenizer.
Waterfall view performance
Traces with 50+ spans previously loaded slowly. The waterfall renderer now virtualizes the row list and handles traces with up to 500 spans without lag. Scroll performance is also improved on mobile.
TypeScript SDK
The TypeScript SDK is now in open beta. Instrument Node.js and Deno LLM applications with the same span model as the Python SDK. Works with Vercel AI SDK, LangChain.js, and direct API clients.
Cost explorer: daily and monthly views
Cost trends can now be viewed by day, week, or month. The anomaly flag highlights any day where spend deviates more than 25% from the rolling 7-day average, making cost spikes easy to spot.
Spanloom early access launch
Spanloom is now available to engineering teams under our early access program. This initial release includes the trace waterfall, basic cost explorer, and Python SDK. Prompt drift detection and RAG audit are in active development.