vocero-s2s/start_pipeline_rag.sh

#!/bin/bash

# ============================================================
# Avvio Pipeline Speech-to-Speech con RAG SERVER-SIDE abilitato
# ============================================================
# - Usa gli stessi endpoint/base URL di start_pipeline.sh
# - Auto-detect del modello LLM via /v1/models
# - Aggiunge tutti i flag --rag_* necessari
# - Persistenza indice NPZ automatica (solo 1a build lenta)
# ============================================================

set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
cd "$SCRIPT_DIR"

# ── Determina interprete Python ──────────────────────────
# Usa ./venv/bin/python come lo script originale start_pipeline.sh
# per essere sicuro di caricare tutte le dipendenze del progetto.
if [ -x "./venv/bin/python" ]; then
  PY="./venv/bin/python"
  echo "✅ Interprete Python: ./venv/bin/python (virtualenv progetto)"
else
  if command -v python3 &>/dev/null; then
    PY="$(command -v python3)"
    echo "⚠️  ./venv/bin/python non trovato, uso python3 di sistema: ${PY}"
    echo "   ⚡ Consigliato: crea il virtualenv con 'python -m venv venv && uv pip install -e .[rag]'"
  else
    echo "❌ Nessun interprete Python trovato (né ./venv/bin/python, né python3)."
    exit 1
  fi
fi

if [ -f ".env" ]; then
  set -a
  . ./.env
  set +a
fi

# ── Endpoint LLM (ereditati da start_pipeline.sh) ──────────
BASE_URL="${LLM_BASE_URL:-http://127.0.0.1:8001/v1}"
DEFAULT_MODEL_NAME="${LLM_MODEL_NAME:-Qwen/Qwen3-32B}"
MODEL_NAME="$DEFAULT_MODEL_NAME"

# ── RAG — Configurazione (personalizza qui) ────────────────
RAG_KB_PATH="${RAG_KB_PATH:-./kb}"            # Cartella documenti + indice
RAG_TOP_K="${RAG_TOP_K:-3}"                    # Chunk max per turno
RAG_THRESHOLD="${RAG_THRESHOLD:-0.30}"         # Soglia similarità coseno [0,1]
RAG_LANGUAGE="${RAG_LANGUAGE:-es}"             # es / it / en (header prompt)
RAG_INJECT_AS="${RAG_INJECT_AS:-system}"       # system (consigliato) / user
RAG_DEVICE="${RAG_DEVICE:-auto}"               # auto / cuda / cpu
RAG_EMBEDDING_MODEL="${RAG_EMBEDDING_MODEL:-paraphrase-multilingual-MiniLM-L12-v2}"
RAG_CHUNK_SIZE="${RAG_CHUNK_SIZE:-512}"
RAG_CHUNK_OVERLAP="${RAG_CHUNK_OVERLAP:-64}"
RAG_FORCE_REBUILD="${RAG_FORCE_REBUILD:-0}"    # 1 = forza ricostruzione indice

# ── Pipeline — porte / repliche ───────────────────────────
WS_HOST="${WS_HOST:-0.0.0.0}"
WS_PORT="${WS_PORT:-12345}"
NUM_PIPELINES="${NUM_PIPELINES:-2}"
LOG_LEVEL="${LOG_LEVEL:-INFO}"

# ── TTS (ereditato da start_pipeline.sh) ──────────────────
TTS_MODEL="${Qwen3_TTS_MODEL:-Qwen/Qwen3-TTS-12Hz-1.7B-Base}"
TTS_REF_AUDIO="${Qwen3_TTS_REF_AUDIO:-female_short.wav}"
TTS_REF_TEXT="${Qwen3_TTS_REF_TEXT:-Chile tiene uno de los acentos más reconocibles del mundo hispano y los generadores de español genérico no logran capturarlo.}"

echo "🎙️ Avvio Pipeline Audio Realtime + RAG (Porta ${WS_PORT})..."
echo "💡 Premi Ctrl+C in questo terminale per fermarla."
echo "🔎 Query del modello LLM in uso su ${BASE_URL}/models..."

DETECTED_MODEL_NAME=$(
  curl -fsS "${BASE_URL}/models" 2>/dev/null | "$PY" -c '
import json
import sys

data = json.load(sys.stdin)
models = data.get("data") or []
first = models[0] if models else {}
print(first.get("id", ""), end="")
' 2>/dev/null || true
)

if [ -n "$DETECTED_MODEL_NAME" ]; then
  MODEL_NAME="$DETECTED_MODEL_NAME"
  echo "🧠 Modello LLM rilevato: $MODEL_NAME"
else
  echo "⚠️ Query /models fallita, uso fallback configurato: $MODEL_NAME"
fi

echo ""
echo "📚 === Configurazione RAG ==="
echo "   KB path        : ${RAG_KB_PATH}"
echo "   Embedding model: ${RAG_EMBEDDING_MODEL}"
echo "   Device         : ${RAG_DEVICE}"
echo "   Top-K          : ${RAG_TOP_K}"
echo "   Soglia         : ${RAG_THRESHOLD}"
echo "   Lingua prompt  : ${RAG_LANGUAGE}"
echo "   Inietta come   : ${RAG_INJECT_AS}"
echo "   Chunk size     : ${RAG_CHUNK_SIZE} (overlap ${RAG_CHUNK_OVERLAP})"
echo "   Force rebuild  : $([[ "$RAG_FORCE_REBUILD" == "1" ]] && echo "SI ⚠️" || echo "no")"
echo "============================"
echo ""
echo "🧩 === Configurazione pipeline ==="
echo "   WS endpoint     : ws://${WS_HOST}:${WS_PORT}"
echo "   Pipeline attive : ${NUM_PIPELINES}"
echo "   LLM backend     : chat-completions"
echo "   TTS model       : ${TTS_MODEL}"
echo "   Log level       : ${LOG_LEVEL}"
echo "=================================="
echo ""
echo "🚀 Avvio tra 2 secondi..."
sleep 2

RAG_FORCE_FLAG=""
if [[ "$RAG_FORCE_REBUILD" == "1" ]]; then
  RAG_FORCE_FLAG="--rag_force_rebuild"
  echo "♻️ RAG FORCE REBUILD attivo — indice verrà ricostruito da zero."
fi

PYTHONPATH=src exec "$PY" src/speech_to_speech/s2s_pipeline.py \
  --mode realtime \
  --ws_host "${WS_HOST}" \
  --ws_port "${WS_PORT}" \
  --num_pipelines "${NUM_PIPELINES}" \
  --log_level "${LOG_LEVEL}" \
  \
  --llm_backend chat-completions \
  --model_name "${MODEL_NAME}" \
  --responses_api_base_url "${BASE_URL}" \
  --responses_api_api_key "EMPTY" \
  \
  --qwen3_tts_backend torch \
  --qwen3_tts_model_name "${TTS_MODEL}" \
  --qwen3_tts_ref_audio "${TTS_REF_AUDIO}" \
  --qwen3_tts_ref_text "${TTS_REF_TEXT}" \
  \
  --rag_enabled \
  --rag_kb_path "${RAG_KB_PATH}" \
  --rag_embedding_model "${RAG_EMBEDDING_MODEL}" \
  --rag_device "${RAG_DEVICE}" \
  --rag_top_k "${RAG_TOP_K}" \
  --rag_threshold "${RAG_THRESHOLD}" \
  --rag_language "${RAG_LANGUAGE}" \
  --rag_inject_as "${RAG_INJECT_AS}" \
  --rag_chunk_size "${RAG_CHUNK_SIZE}" \
  --rag_chunk_overlap "${RAG_CHUNK_OVERLAP}" \
  ${RAG_FORCE_FLAG}