What's new

Changelog.

Every change to the SEER API and dashboard, documented plainly. No surprises. We never make breaking changes within a major version.

v1.4.0Minor14 November 2025

The prescribe engine got a major upgrade this release. Recommendations are now more specific, ranked by financial impact, and include an estimated time to fix. We also added native CrewAI support and a new cost spike webhook event.

Added
Prescribe engine v2. Recommendations are now ranked by estimated dollar impact, not just severity. Each recommendation includes a "time to fix" estimate so you can prioritise by effort too.
Native CrewAI integration. Attach SEER as a CrewAI callback handler and it observes every agent step automatically. No per-call wrapping needed. See the API docs for setup.
cost.spike webhook event. SEER now fires a webhook when cost per call doubles compared to your 7-day average for a context. Configure the sensitivity multiplier in your dashboard.
Session-level quality scores. When you pass a seer_session_id, SEER now calculates and returns a quality score for the full session — not just individual calls. Useful for evaluating multi-turn conversations end to end.
Fixed
Prompt drift detection false positives. Whitespace changes in prompts were incorrectly triggering drift alerts. SEER now normalises whitespace before comparing, so you only get alerts for meaningful changes.
Dashboard latency chart timezone issue. Charts were displaying latency data in UTC regardless of your account timezone setting. This is now consistent with the timezone you have configured.
Changed
Quality score calculation. The groundedness dimension now accounts for factual consistency across a session, not just a single call. Scores on multi-turn contexts may shift slightly as a result. Single-call contexts are unaffected.
v1.3.2Patch28 October 2025

Reliability patch. Fixed two edge cases in the anomaly detector that were causing missed alerts under specific load patterns, and resolved a memory issue in the Node.js SDK that appeared after many hours of continuous use.

Fixed
Anomaly detector missed alerts at high concurrency. When more than 15 parallel calls came in at once, the anomaly detector's baseline comparison could miss spikes. Fixed by making the comparison lock-free.
Anomaly detector triggered on first call. The first call in a new context had no baseline to compare against, which was sometimes triggering a false anomaly. SEER now waits for at least 10 calls before anomaly detection activates for a new context.
Node.js SDK memory growth. The SDK was retaining trace references in memory indefinitely in long-running processes. References are now released after the response is returned.
evaluate() returned wrong dimension labels. A label mapping bug caused the groundedness and task_completion keys to swap in the breakdown object. The scores themselves were always correct. Now the labels match.
v1.3.0Minor7 October 2025

This release focuses on teams building with multiple AI providers. SEER now supports AWS Bedrock and Azure OpenAI natively, and introduces cross-provider comparison reporting so you can see which provider is performing better for each context.

Added
AWS Bedrock support. Pass your Bedrock client and model ARN to seer.observe() the same way you would with OpenAI. Claude, Titan, Llama, and Mistral on Bedrock are all supported.
Azure OpenAI support. Works with your Azure endpoint and deployment name. SEER handles the credential format differences automatically.
Cross-provider comparison reports. In your dashboard, you can now run a side-by-side report for one context across multiple providers. Shows quality, cost, and latency differences with a clear recommendation on which provider is working best for that use case.
API key scoping. You can now restrict a key to specific contexts. A key scoped to "customer-support" cannot record or read data from any other context. Useful for giving different teams access to only their own data.
Webhook delivery retry. If your webhook endpoint returns a non-2xx response, SEER retries delivery up to 5 times with exponential backoff over 24 hours. You can view retry history and failed deliveries in your dashboard.
Fixed
Cost calculation for streaming responses. When using streaming mode, token counts were being under-counted because the final usage object was arriving after SEER's recording window. SEER now waits for the stream to fully complete before computing cost.
Changed
recommend() response shape. The impact field now returns an object with impact.description (plain English) and impact.estimated_saving_usd (number, monthly estimate) instead of a plain string. Update your code if you read this field programmatically.
v1.2.0Minor15 September 2025

The evaluate() method graduates from beta. Also adds LangChain and LlamaIndex native integrations, plus a new dashboard view that lets you track quality trends over time without writing any queries.

Added
evaluate() is now stable. Exits beta. The response shape is now stable and covered by our versioning guarantee. All beta users are automatically migrated — no code changes needed.
LangChain callback handler. Import SeerCallbackHandler from the Python SDK and pass it to any LangChain chain. SEER observes every LLM call in the chain without you wrapping each one individually.
LlamaIndex observer. Same pattern as LangChain — attach once at the index level, SEER records everything.
Quality trend chart in dashboard. A new time-series chart on every context page shows how quality score has moved over the past 7, 30, or 90 days. Regression events are marked on the timeline.
Daily digest email. Optional. Subscribe in your dashboard to receive a daily email at 8am with your top 3 recommended actions, yesterday's quality summary, and any unresolved anomalies.
Fixed
Intermittent 503 on high-volume accounts. Accounts sending more than 200 calls per second were occasionally getting 503 errors during peak hours. Resolved by adding additional capacity to the ingestion layer.
v1.1.0Minor22 August 2025

Webhooks ship in this release, along with the recommend() method and the first version of the SEER dashboard. This is the release where SEER became a complete product rather than just an API.

Added
Webhooks. Subscribe to anomaly, quality, eval, cost, and plan events. SEER sends a signed POST request to your endpoint in real time when an event fires.
recommend() method. Call seer.recommend(context="...") and get back a ranked list of specific actions to improve that context. Each action includes a plain English explanation and an estimated impact.
SEER dashboard (beta). A web interface showing call history, quality trends, anomaly timeline, and recommendations per context. Available to all accounts at dashboard.seer.ai.
Slack integration. Connect your Slack workspace from the dashboard. SEER posts a message to your chosen channel when an anomaly is detected or quality drops below threshold.
v1.0.0Major1 August 2025

Initial public release. SEER launches with observe(), evaluate() in beta, trace(), OpenAI and Anthropic support, and the core intelligence layer. This is the one API that replaces four to six separate monitoring tools.

Added
observe() method. Wrap any model call. SEER records it, scores it, and returns quality, cost, latency, anomaly, and prompt drift data alongside your normal response.
evaluate() method (beta). Run structured quality checks against a rubric. Returns a pass/fail result, a score, and a plain English explanation of any failure.
trace() method. Retrieve the full record of any past call by its ID. Every record is immutable — it cannot be altered after the fact.
OpenAI support. All GPT-4o, GPT-4 turbo, and GPT-3.5 models. Streaming and non-streaming.
Anthropic support. Claude 3.5 Sonnet, Claude 3 Opus, and Claude 3 Haiku. Streaming and non-streaming.
Python SDK. pip install cadence-seer.
Node.js SDK. npm install @cadence/seer.
REST API. Call SEER directly from any language using standard HTTP.