fix: route conversational queries early in stream_answer to bypass RAG retrieval and ensure correct greeting/identity responses

feat: enhance conversational query prompts and fallback handlers to support Risu Solutions / BAMS curriculum identity

fix: copy all .so dynamic libraries to /usr/local/bin to resolve llama-server backend scanning path

fix: copy all .so dynamic libraries from llama.cpp tarball to /usr/local/lib and run ldconfig

debug: run llama-server logging to file, do not block uvicorn, expose server log in debug endpoint

fix: extract pre-built ChromaDB index zip on startup to bypass slow CPU indexing

fix: install libgomp1 in Dockerfile to satisfy llama-server OpenMP shared library dependency

debug: add /debug/llm endpoint to diagnose completions traceback

fix: deploy native llama-server sidecar from pre-built tarball (supports n_parallel, fixes fallback mode)

tune: raise context budget 700->1400 tokens + max_tokens 1200->1800 for 600-800 word answers

fix: remove sidecar llama-server - revert to direct uvicorn startup (was working in 10/10 test)

fix: use python -m llama_cpp.server with --n_parallel 3 (no binary download needed, pip install already covers it)

fix: download pre-built llama-server binary instead of compiling from source - reduces build time from 20min to under 2min

feat: complete BAMS syllabus database ingestion & local LLM config

feat: complete BAMS syllabus database ingestion & local LLM config

feat: optimize deployment payload using zipped database index