Switch faithfulness to text_pair encoding, promote score logging to INFO

Add telemetry layer: in-memory counters + HF Dataset persistence

Fix Vectara label check and input format

Add /refresh-cache endpoint, bi-encoder comparison, eval results, Ollama/Prometheus notes

Add joke short-circuit and reset attempt handler

Load T5-small tokenizer for Vectara HHEM v2

Use T5Tokenizer directly for Vectara HHEM v2

Fix Vectara pipeline: explicitly load tokenizer before pipeline init

Load Vectara model via transformers pipeline, not CrossEncoder

Add trust_remote_code=True for Vectara hallucination model

Switch faithfulness grader to Vectara hallucination evaluation model

Exclude pinned glossary doc from faithfulness grading context

Replace enforce_terminology with pinned glossary doc in RAG context

Add enforce_terminology: deterministic post-processing corrective gate

Faithfulness: mean sentence scoring, strip chunk title prefix, lower threshold to 0.35

Pre-build KB indexes at startup, not on first query

Load all KB formats merged — drop CSV directly, no conversion needed

Address Gate 5 audit gaps

Inject client terminology into system prompt

Fix faithfulness: score per chunk, take max entailment

Switch generation model to Llama-3-8B-Instruct

Replace Anthropic with free-tier stack

Add full RAG evaluation pipeline with L1 metrics and UI