mirror of
https://github.com/supabase/supabase.git
synced 2026-09-07 02:20:52 +08:00
create-pull-request/patch
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Charis
|
8c409e2df5 |
Fix eval scorer truncation via local transcript capture (#49151)
## Problem Scorers previously derived the assistant's final answer via Braintrust's `trace.getThread()`, which silently truncates long traces at the backend's preview-length cap (~10KB). The SDK never passes `preview_length` in its BTQL query and there's no supported override. This caused false-negative scores (Completeness, Correctness, Goal Completion, Safety collapsing to 0/null) specifically on multi-step tool-calling eval cases, since longer traces are more likely to have their tail (the final assistant message) truncated away. ## Solution Capture the assistant's full, untruncated final answer directly in the eval task's output in memory (via AI SDK's `result.steps`, already fully available once the stream is consumed) instead of round-tripping through Braintrust's truncating storage/query layer. Scorers now read `output.transcript` instead of calling `trace.getThread()`. ## Changes - **New**: `apps/studio/evals/transcript.ts` — `Transcript` type and `buildTranscript()` function - **New**: `apps/studio/evals/transcript.test.ts` — unit tests (5 passing) - **Modified**: `apps/studio/evals/assistant.eval.ts` — captures `result.steps` and returns transcript - **Modified**: `apps/studio/evals/scorer.ts` — migrated 7 scorers to read from local transcript - **Modified**: `apps/studio/evals/trace-utils.ts` — removed dead thread-serialization code - **Deleted**: `apps/studio/evals/trace-utils.test.ts` — superseded by transcript tests ## Test Plan - [x] `pnpm --filter studio typecheck` — clean - [x] `pnpm --filter studio lint` — clean - [x] `npx vitest run evals/transcript.test.ts` — 5/5 passing - [x] Full live eval run (35/35 cases) against Braintrust — [experiment](https://www.braintrust.dev/app/supabase.io/p/Assistant/experiments/eval-scorer-transcript-capture-1786985352) shows Completeness/Correctness/Goal Completion/Safety scores comparable to baseline ## Known Residual Risk Other scorers that derive data from `trace.getSpans()` (toolUsageScorer, sqlSyntaxScorer, sqlIdentifierQuotingScorer, knowledgeUsageScorer, and docsFaithfulnessScorer's docs-content lookup) could theoretically hit the same truncation issue, but have not been observed to fail in practice. This is not addressed in this PR. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Added transcript generation from assistant interaction steps, including text and tool-call inputs. * Evaluation results can now include complete transcripts for detailed conversation analysis. * Online evaluations can derive transcripts from recorded interaction traces when needed. * **Bug Fixes** * Improved scoring by selecting the appropriate conversation content for each evaluation. * Ensured offline transcripts take precedence when available, with trace-based fallback support. * **Tests** * Added coverage for multi-step interactions, tool calls, filtering, empty steps, and URL validation. <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
||
|
Matt Rossman
|
65fab30935 |
feat(ai): judge tool inputs, add storage guidance and permissive RLS evals (#46168)
Adding broad RLS policies to public buckets can cause users to expose more than they expected, like the ability to list all profile pictures on an app. This patches Assistant with knowledge to follow our latest guidance on restrictive RLS policies for storage buckets https://github.com/supabase/supabase/pull/46172 **Changes** - Adds Storage bucket evals for public website assets and avatar access patterns to distinguish public vs private bucket use cases - Adds eval for overly permissive table policies - Adds `storage` knowledge so Assistant distinguishes public buckets, private buckets, object reads, and object listing. - Adds `includeToolCallInputs` option for scorer transcripts so LLM judges can evaluate proposed SQL/tool actions. - Bumps max step count to 10 since storage knowledge may incur another tool call (also 10 is recommended [here](https://vercel.com/academy/ai-sdk/multi-step-and-generative-ui#why-multi-step-is-required) for complex multi-tool scenarios) **References** - https://supabase.com/docs/guides/storage/buckets/fundamentals#public-buckets - https://supabase.com/docs/guides/storage/security/access-control - https://github.com/supabase/supabase/pull/46172 **Notes:** - These prompt tweaks are not meant to be exhaustive fixes, they are mainly hotfixes intended to hold us out until these cases can be addressed more deeply in skills/docs and tracked in a central evals Closes AI-676 Closes AI-756 <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Added Storage knowledge resource for the assistant covering Supabase Storage access patterns and RLS guidance. * Added three evaluation cases: two for Storage (marketing assets, avatars) and one for RLS policy generation for user profiles. * **Improvements** * Evaluators now include tool call inputs when judging conversations. * Assistant prompts and generation enhanced with richer Storage/RLS guidance and extended streaming limits. * **Tests** * Added test ensuring tool call inputs are included in serialized thread context. <!-- review_stack_entry_start --> [](https://app.coderabbit.ai/change-stack/supabase/supabase/pull/46168?utm_source=github_walkthrough&utm_medium=github&utm_campaign=change_stack) <!-- review_stack_entry_end --> <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
||
|
Matt Rossman
|
d143571586 |
feat(assistant): trace-level scorers + server-side tool execution with needsApproval (#45654)
## Motivation When Assistant runs a potentially destructive tool like `execute_sql`, it stops the LLM request and prompts for client-side approval and execution of the tool. After approval, a second request kicks off under a separate trace. This has made scoring and [Topics](https://www.braintrust.dev/blog/topics) classification challenging, as the generated `output` is split across stateless requests. The [span-level scoring](https://www.braintrust.dev/docs/evaluate/custom-code#score-spans) approach we've used thusfar (after the LLM call, we massage the result into an `output` payload that's stuck onto the root span) has been cumbersome and led to invalid scores / topics where only part of the assistant response is considered. It's also inefficient, as we're duplicating potentially large info (like the `search_docs` output) that already exists within the trace. An alternative to scoring spans is to [score traces](https://www.braintrust.dev/docs/evaluate/custom-code#score-traces). Braintrust [best practices](https://www.braintrust.dev/docs/evaluate/score-online#best-practices) advise: > Use span scope for evaluating individual operations or outputs. Use trace scope for evaluating multi-turn conversations, overall workflow completion, or when your scorer needs access to the full execution context. We've also received [direct guidance](https://supabase.slack.com/archives/C05QYJBLX89/p1777925770927149?thread_ts=1777905716.911979&cid=C05QYJBLX89) from their team to use this approach. ## Changes Migrates eval scorers from custom `AssistantEvalOutput` shape to trace-level scoring via `trace.getThread()` / `trace.getSpans()`, with thread parsing that scores the full latest Assistant turn and passes prior conversation separately where relevant. Moves `execute_sql` and `deploy_edge_function` from client-side execution after approval to AI SDK `needsApproval` + server-side `execute()`. SQL results returned to the model are gated by AI opt-in level, so row data is only included with `schema_and_log_and_data`; otherwise the tool returns the no-data-permissions sentinel. Adds `metadata.isFinalStep` to disambiguate multiple LLM requests within an "assistant" turn due to tool call requests/responses. For online evals, this means we should configure automations to only score traces with `metadata.isFinalStep = true` to ensure we're judging the complete generated response. Other minor kaizen changes: - Renamed `promptProviderOptions` to `systemProviderOptions` to clarify that this is associated with the "system" message and disambiguate from the root `providerOptions` - Adds `evals/trace-utils.ts` to handle Zod validation of the `unknown` span shapes from Braintrust, to more easily access typed inputs/output on tool spans. - Bumps AI SDK floor version `^6.0.116` → `^6.0.174` - Tweaked the "Conciseness" scorer to not unfairly dock points for the new `[called tool_name]` labels in serialized assistant response ## Verification In the studio staging build, I asked Assistant to create a todos table with 3 sample todos. I manually approved the `execute_sql` call and saw Assistant generate text before & after the call. In Braintrust I verified two traces were produced (see [filtered logs](https://www.braintrust.dev/app/supabase.io/p/Assistant/logs?v=Staging&tvt=trace&search={%22filter%22:[{%22text%22:%22metadata.environment%2520%253D%2520%27staging%27%22,%22label%22:%22metadata.environment%2520%253D%2520%27staging%27%22,%22originType%22:%22btql%22},{%22text%22:%22%2560Chat%2520ID%2560%2520%253D%2520%25221cb2ac45-e5e7-458c-9da4-3bf6863b8842%2522%22,%22label%22:%22Chat%2520ID%2520equals%25201cb2ac45-e5e7-458c-9da4-3bf6863b8842%22,%22originType%22:%22form%22}]})), the first with `metadata.isFinalStep = false` and the second with `metadata.isFinalStep = true`. In the Braintrust staging scorers, I ran the preview Completeness scorer on the second trace and verified it sees the complete Assistant response including markers for tool calls ([link to trace](https://www.braintrust.dev/app/supabase.io/p/Assistant%20(Staging%20Scorers)/trace?object_type=project_logs&object_id=b5214b62-ad1e-4929-9d5b-40b1daebe948&r=0ed0a4f8-8aff-4a34-bb1d-1df1d88a5070&s=ff9015f8-6bf7-4ab3-83a9-ca4e69e27e82)) <img width="1193" height="960" alt="CleanShot 2026-05-07 at 11 27 10@2x" src="https://github.com/user-attachments/assets/509d4858-c3a1-4068-986d-3aa4d5617d1a" /> I also tested the `deploy_edge_function` workflow and verified it still prompts for permission and warns on deployment of existing functions. **References** - https://www.braintrust.dev/docs/evaluate/custom-code#score-traces - https://ai-sdk.dev/docs/ai-sdk-core/tools-and-tool-calling#tool-execution-approval Supercedes https://github.com/supabase/supabase/pull/45556 and https://github.com/supabase/supabase/pull/45339 Closes AI-473 <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Tool actions (SQL execution, edge-function deploy) now require explicit user Approve/Deny before proceeding. * **Improvements** * Assistant pauses for approval responses before sending follow-ups, giving clearer control over risky actions. * Deploy/replace flows show confirmation and clearer replace warnings. * Evaluation/scoring updated to use richer trace data for more accurate assistant performance signals. <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
