check_embedding_drift
Pin and re-check a 16-string canary against the active embedding provider. Catches silent provider model swaps (OpenAI/Voyage/etc.) that quietly degrade hybrid retrieval. First call (or with capture=true) saves the baseline; subsequent calls report max cosine distance vs baseline. Read-only or wr...
This record as markdown: /tools/nikolai-vysotskyi-trace-mcp/check-embedding-drift.md
What check_embedding_drift does on Trace
AI agents call check_embedding_drift to retrieve information from Trace without modifying anything. It is typically the context-gathering step in research, monitoring, and reporting workflows, before the agent takes action elsewhere.
Why check_embedding_drift is rated Low
While the tool has a write component (capture=true saves baseline), its core purpose is monitoring and diagnostics—reading embedding data and comparing it to detect drift. The write operation is administrative baseline-setting, not a business-logic mutation. The 'read-only' framing in the description emphasizes the default behavior. No code execution, deletion, or external effects triggered by the results.
From the tool's definition Tool performs 'pin and re-check' with 'read-only or write-only (capture)' modes. The capture mode saves a baseline (write operation), but the primary function is checking/comparing embeddings against a baseline—a retrieval and analysis task.
Attacks that exploit this kind of access
The rule that runs check_embedding_drift safely
PolicyLayer is an MCP gateway: it sits between your AI agents and Trace, and checks every tool call against a rule you set before the call runs. Nothing changes on the server itself. For check_embedding_drift, this is the rule to start with:
check_embedding_drift is read-only, so it stays allowed. Everything else on the server is denied unless you say otherwise.
The button opens the PolicyLayer dashboard: create your workspace, connect Trace, apply this rule, and every check_embedding_drift call is checked against it from then on.
Questions about check_embedding_drift
Pin and re-check a 16-string canary against the active embedding provider. Catches silent provider model swaps (OpenAI/Voyage/etc.) that quietly degrade hybrid retrieval. First call (or with capture=true) saves the baseline; subsequent calls report max cosine distance vs baseline. Read-only or write-only (capture). Returns JSON: { status, message, max_distance?, mean_distance?, per_string? }. It is categorised as a Read tool in the Trace MCP Server, which means it retrieves data without modifying state.
Register the Trace MCP server in PolicyLayer and add a rule for check_embedding_drift: allow, deny, rate-limit, or require approval. Point your MCP client at the PolicyLayer proxy URL and the rule is enforced on every call, before it reaches Trace. Nothing to install.
check_embedding_drift is a Read tool with low risk. Read-only tools are generally safe to allow by default.
Yes. Add a rate_limit block to the check_embedding_drift rule in your PolicyLayer policy. For example, setting max: 10 and window: 60 limits the tool to 10 calls per minute. Rate limits are tracked per agent session and reset automatically.
Set action: deny in the PolicyLayer policy for check_embedding_drift. The AI agent will receive a policy violation error and cannot call the tool. You can also include a reason field to explain why the tool is blocked.
check_embedding_drift is provided by the Trace MCP server (nikolai-vysotskyi/trace-mcp). PolicyLayer sits as a proxy in front of this server to enforce policies before tool calls reach the server.
More on Trace, and thousands of servers like it.
This server
Across the catalogue