Repository navigation
feat(chat): reply-grounding check (log only), 62-case eval harness, content-free turn log and health endpoint - #230
Merged
Conversation
…d stream routes with plan, single log line
|
Deployment failed for project okara-alternative with the following error: Learn More: https://vercel.com/kumar2421s-projects?upgradeToPro=build-rate-limit |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Keeps the AI CMO honest and observable. checkReplyGrounding flags numbers/percents/dates/domains in a reply that appear in no tool result (LOG-ONLY: never changes or blocks a reply; accepts 1,234 vs 1234, k shorthand, CTR 0.034 as 3.4%, ignores list numbering/durations/the current year). 62 eval cases (fix-first, traffic, did-my-fix-work, site health, never-audited vs clean, competitors, leads, GEO, out-of-scope, prompt injection, tool failure, plan gating) with a deterministic CI runner that validates the PIPELINE (budgets, tool-result wrapping and caps, unknown/withheld tools never run, never-audited vs clean reaches the model, grounding flags invented numbers) - not the real model's judgement; gating cases use the REAL entitlement port (free withholds get_geo_status and get_audit_summary; leads are not locked). Manual live script scripts/assistant-eval.mts (refuses without ASSISTANT_EVAL_API_KEY and in CI; never run live yet; --dry works). Observability: ONE structured JSON line per turn on BOTH /api/chat and /api/chat/stream (evt, ok, provider, model, candidateIndex, fallback, attempts, tools, toolMs, totalMs, errorKind, groundingFlags, plan, hashed user id u) - never message text, tool results, emails, thread ids or deltas (leak tests); legacy text line silenced; in-memory counters (turns, errors by kind, fallbacks, p50/p95 over last 200); GET /api/admin/assistant-health behind the CRON_SECRET bearer pattern (per instance). Docs in docs/assistant-evaluation.md and docs/assistant-observability.md. 2226 tests pass. Caveat: groundingFlags is 0 for real-token-streamed replies (only tool-path/non-streamed replies are checked).
🤖 Generated with Claude Code