code/src/generators/rag_config.py

"""
Centralized configuration for the RAG system.
Edit this file to update paths, model names, and other settings in one place.
"""

import os
import torch
from pathlib import Path

# Base data directory (edit as needed)
BASE_DATA_DIR = os.path.abspath(os.path.join(os.path.dirname(__file__), '../../data/knowledge_base'))

# ── Qdrant ────────────────────────────────────────────────────────────────────
QDRANT_URL = os.getenv("QDRANT_URL", "http://localhost:6333")
QDRANT_API_KEY = os.getenv("QDRANT_API_KEY", None)

QDRANT_COLLECTIONS = {
    "cve":   "cve_chunks",
    "mitre": "mitre_techniques",
    "capec": "capec_patterns",
    "cwe":   "cwe_entries",
}

DENSE_VECTOR_NAME  = "dense"
SPARSE_VECTOR_NAME = "sparse"

# ── BM25 vocabulary ───────────────────────────────────────────────────────────
BM25_VOCAB_PATH = os.path.join(BASE_DATA_DIR, "bm25", "vocab.json")
BM25_K1 = 1.5
BM25_B  = 0.75
BM25_MIN_DF = 2       # drop tokens with doc-frequency < this
BM25_MAX_VOCAB = 200_000

# ── Neo4j ─────────────────────────────────────────────────────────────────────
NEO4J_URI      = os.getenv("NEO4J_URI",      "bolt://localhost:7687")
NEO4J_USER     = os.getenv("NEO4J_USER",     "neo4j")
NEO4J_PASSWORD = os.getenv("NEO4J_PASSWORD", os.getenv("NEO4J_AUTH", "neo4j/password").split("/")[-1])

# ── RAG export paths ──────────────────────────────────────────────────────────
RAG_EXPORTS_DIR = os.path.join(BASE_DATA_DIR, "rag_exports")

CVE_CHUNKS_PATH   = os.path.join(RAG_EXPORTS_DIR, "cve_chunks.json")
MITRE_CHUNKS_PATH = os.path.join(RAG_EXPORTS_DIR, "mitre_chunks.json")
CAPEC_CHUNKS_PATH = os.path.join(RAG_EXPORTS_DIR, "capec_chunks.json")
CWE_CHUNKS_PATH   = os.path.join(RAG_EXPORTS_DIR, "cwe_chunks.json")

# Legacy alias kept for any code that still imports it
CVE_DATA_PATH = CVE_CHUNKS_PATH

# Year-based data paths (used by search_cves_by_year fallback)
CVE_YEAR_PATHS = {
    str(yr): os.path.join(BASE_DATA_DIR, f"enhanced_documents_cve_{yr}.json")
    for yr in range(1999, 2027)
}

# ── Embedding model ───────────────────────────────────────────────────────────
EMBEDDING_MODEL_NAME = "microsoft/harrier-oss-v1-270m"  # 640-dim, MTEB 66.5
DENSE_VECTOR_SIZE    = 640

# ── GPU / compute ─────────────────────────────────────────────────────────────
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
EMBEDDING_BATCH_SIZE = 256  # sweet spot for harrier on 4GB VRAM
MAX_SEQ_LENGTH = 1024       # fits full CVE description + severity sections

# ── Chunking parameters (kept for reference; real chunking is in chunkers.py) ─
CHUNK_SIZE    = 512
CHUNK_OVERLAP = 50

# ── Search parameters ─────────────────────────────────────────────────────────
VECTOR_SEARCH_TOP_K = 50
RERANK_TOP_K        = 10
HYBRID_ALPHA        = 0.7   # weight for vector vs keyword (used if manual fusion needed)

# ── Performance ───────────────────────────────────────────────────────────────
MAX_QUERY_LENGTH    = 1000
MAX_CONTEXT_LENGTH  = 8000
MAX_RESPONSE_LENGTH = 2000
MAX_SEARCH_RESULTS  = 50

# ── Timeouts (seconds) ────────────────────────────────────────────────────────
EMBEDDING_TIMEOUT     = 5
SEARCH_TIMEOUT        = 10
LLM_GENERATION_TIMEOUT= 30

# ── API rate limiting ─────────────────────────────────────────────────────────
MAX_REQUESTS_PER_MINUTE = 60
RATE_LIMIT_WINDOW       = 60

# ── LLM (OpenAI-compatible API) ────────────────────────────────────────────────
# Uses the llms/ package. Configure via env vars or edit directly.
# Works with any OpenAI-compatible endpoint: Ollama, vLLM, LM Studio, OpenAI, etc.

LLM_ENABLED = os.getenv("LLM_ENABLED", "false").lower() == "true"
LLM_PROVIDER = os.getenv("LLM_PROVIDER", "openai")  # "openai" | "ollama" | "vllm"
LLM_BASE_URL = os.getenv("LLM_BASE_URL", "http://localhost:11434/v1")
LLM_MODEL_NAME = os.getenv("LLM_MODEL_NAME", "llama3.1:8b-instruct-fp16")
LLM_API_KEY = os.getenv("LLM_API_KEY", os.getenv("OPENAI_API_KEY", None))
LLM_TEMPERATURE = float(os.getenv("LLM_TEMPERATURE", "0.1"))
LLM_MAX_TOKENS = int(os.getenv("LLM_MAX_TOKENS", "800"))
LLM_ENABLE_REASONING = os.getenv("LLM_ENABLE_REASONING", "false").lower() == "true"

def get_llm_config() -> dict:
    """Return a config dict compatible with llms.factory.LLMFactory.create_from_config()."""
    return {
        "enabled":           LLM_ENABLED,
        "provider":          LLM_PROVIDER,
        "model_name":        LLM_MODEL_NAME,
        "base_url":          LLM_BASE_URL,
        "api_key":           LLM_API_KEY,
        "default_temperature": LLM_TEMPERATURE,
        "default_max_tokens":  LLM_MAX_TOKENS,
        "enable_reasoning":  LLM_ENABLE_REASONING,
    }

# Legacy aliases kept for any old code that still imports them
def get_hf_token():
    token_file = Path(__file__).parent.parent.parent / "llama_token.txt"
    if token_file.exists():
        with open(token_file) as f:
            return f.read().strip()
    return os.getenv("HF_TOKEN", "")

HF_TOKEN = get_hf_token()

USE_OLLAMA          = os.getenv("USE_OLLAMA", "false").lower() == "true"
OLLAMA_HOST         = os.getenv("OLLAMA_HOST", "http://localhost:11434")
OLLAMA_MODEL_PRIMARY= os.getenv("OLLAMA_MODEL_PRIMARY", "llama3.1:8b-instruct-fp16")
OLLAMA_MODEL_FAST   = os.getenv("OLLAMA_MODEL_FAST",    "llama3.1:8b-instruct-fp16")

HF_MODEL_PRIMARY = "meta-llama/Llama-3.1-8B-Instruct"
HF_MODEL_FAST    = "meta-llama/Llama-3.1-8B-Instruct"

# ── API server ────────────────────────────────────────────────────────────────
API_HOST    = os.getenv("API_HOST",    "0.0.0.0")
API_PORT    = int(os.getenv("API_PORT",    "8000"))
API_WORKERS = int(os.getenv("API_WORKERS", "1"))

# ── Logging ───────────────────────────────────────────────────────────────────
LOGGING_LEVEL = os.getenv("RAG_LOGGING_LEVEL", "INFO")