Add eval harness

Initial deploy from GitHub main

Add meta-word avoidance rule to default security prompt

Fix false escalate: text-based detection only

Fix false escalate check on no_context replies

Eval: out_of_scope is not escalation

Out-of-scope reply when RAG has no context

Docs: grounding off by default; claim-level path

Disable whole-answer grounding guardrails by default

Docs: id-only scenarios; ENABLE_TOOLS gates eval tools

Drop scenario kind/topic; id-only identification

Docs: evals scenario kinds + tool access vs check

Simplify scenario kinds; clarify tool access vs check

Move eval admin UI into evals/ui.py (app-agnostic)

Simplify eval scenario editor for non-technical users

Fix nested-expander crash in eval admin panel

Copy evals/ into Docker image so the eval admin panel can import it

Add in-app eval harness with Google Sheet-backed scenarios

Pass the full dialog into answer context

Update README to match current RAG pipeline

Fix collapsed rerank scores: larger chunks and faithful query rewrite

Add social-intent classifier; disable moderation

Point ExecChat knowledge base to the new Google Doc.

Align INVITE_PATTERNS with OnboardingChat

Require explicit submit instead of firing on every rerun

Point knowledge base to the new Google Doc

Drop KEEP_TOP_K; select context by rerank threshold only

Re-introduce 0.1 rerank threshold alongside top-K

Fix nested-expander crash in debug view

Expose grounding inputs in debug view

Make query rewriter preserve intent and drop boilerplate

Improve reranker recall: query rewrite, relative top-K, smaller chunks

Re-enable grounding as a hard gate with the softened prompt.

Show all rerank candidates with pass/fail in debug for retrieval calibration.

Soften grounding check and make it advisory to stop false-positive escalations.

Add output guardrails: moderation, format checks, grounding via gpt-4.1-nano.

Show rerank score threshold in model settings caption.

Lower rerank score threshold to 0.1.

Hardcode rerank score threshold instead of reading from env.

Set default rerank score threshold to 0.15.

Add retrieval relevance guardrail and prompt safety rules.

Increase reranked RAG chunks from 5 to 7 per response.

Remove empty rerank score from RAG debug UI.

Reduce reranked RAG chunks from 7 to 5 per response.

Improve RAG retrieval with smaller chunks, BGE reranker, and score debug.

Switch to GPT-4.1 mini and reduce context token usage

Load system prompt from Google Doc and add temperature controls.

ExecChat: app_2, tools, Docker (HF-compatible, no binary files)