RAG1 / config.py
wassim2433's picture
Switch embedding model to Alibaba-NLP/gte-Qwen2-7B-instruct via AI Grid API
c080f15
Raw
History Blame Contribute Delete
1.82 kB
import os
# ── API Keys ──────────────────────────────────────────────
AIGRID_API_KEY_GPT = os.getenv("AIGRID_API_KEY_GPT", "sk-XaDrxkmJNvrp04SfkHT2ig")
AIGRID_API_KEY_GEMMA = os.getenv("AIGRID_API_KEY_GEMMA", "sk-eQYZ67KgWjIMcPZb6SwwKg")
AIGRID_API_KEY_EMBEDDING = os.getenv("AIGRID_API_KEY_EMBEDDING", "sk-CNXeCrle_NG4iCre1XvcAA")
OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY", "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")
# ── API Bases ─────────────────────────────────────────────
AIGRID_API_BASE = "http://app.ai-grid.io:4000/v1"
OPENROUTER_API_BASE = "https://openrouter.ai/api/v1"
# ── Model Settings ────────────────────────────────────────
LLM_MODEL = "gpt-oss-120b"
EMBEDDING_MODEL = "Alibaba-NLP/gte-Qwen2-7B-instruct"
LLM_TEMPERATURE = 0.1 # Low = less hallucination
MAX_TOKENS = 1024
# ── Chunking Settings ─────────────────────────────────────
CHUNK_SIZE = 500
CHUNK_OVERLAP = 100
# ── Retrieval Settings ────────────────────────────────────
TOP_K = 5
SIMILARITY_CUTOFF = 0.3 # Below this = "not found"
# ── Paths ─────────────────────────────────────────────────
DOCUMENTS_DIR = "./documents"
FAISS_INDEX_PATH = "/data/faiss_index"