Datasets:

DuyTa
/

cve-kgrag-db

"""
Central configuration for the Agentic RAG system.
Edit AGENT_* env vars to tune behavior without touching code.
All prompts live in prompts.py (single source of truth).
"""

import os

# ── Agent loop ──────────────────────────────────────────────────────────────────
AGENT_MAX_RETRIES = int(os.getenv("AGENT_MAX_RETRIES", "2"))
AGENT_SEARCH_TIMEOUT = float(os.getenv("AGENT_SEARCH_TIMEOUT", "15.0"))
AGENT_GENERATION_TIMEOUT = float(os.getenv("AGENT_GENERATION_TIMEOUT", "180.0"))
AGENT_MAX_CONTEXT_TOKENS = int(os.getenv("AGENT_MAX_CONTEXT_TOKENS", "4000"))
AGENT_DEFAULT_K = int(os.getenv("AGENT_DEFAULT_K", "10"))

# ── Mem0 (self-hosted, OSS) ───────────────────────────────────────────────────
MEM0_ENABLED = os.getenv("MEM0_ENABLED", "true").lower() == "true"
MEM0_USER_ID_DEFAULT = os.getenv("MEM0_USER_ID_DEFAULT", "default_session")
MEM0_MAX_MEMORIES = int(os.getenv("MEM0_MAX_MEMORIES", "5"))
MEM0_COLLECTION_NAME = os.getenv("MEM0_COLLECTION_NAME", "mem0_cve_memories")

# ── Web search (optional, requires Tavily API key) ─────────────────────────────
WEB_SEARCH_ENABLED = os.getenv("WEB_SEARCH_ENABLED", "false").lower() == "true"

# ── Grading backend ────────────────────────────────────────────────────────────
# "reranker" — use Jina Reranker v3 (fast, no tokens consumed, recommended)
# "llm"      — use batched LLM call (slower, costs tokens, fallback)
GRADE_BACKEND = os.getenv("GRADE_BACKEND", "reranker")

# ── Reranker (Jina Reranker v3) ────────────────────────────────────────────────
RERANKER_BACKEND = os.getenv("RERANKER_BACKEND", "local")   # "local" | "api"
RERANKER_MODEL = os.getenv("RERANKER_MODEL", "jinaai/jina-reranker-v3")
RERANKER_THRESHOLD = float(os.getenv("RERANKER_THRESHOLD", "0.5"))  # sigmoid-normalized; 0.5 = neutral

# ── Grading thresholds (LLM path) ──────────────────────────────────────────────
GRADE_RELEVANCE_THRESHOLD = float(os.getenv("GRADE_RELEVANCE_THRESHOLD", "1"))  # min relevant docs
GRADE_MIN_PASSING_SCORE = float(os.getenv("GRADE_MIN_PASSING_SCORE", "0.5"))

# ── Streaming ──────────────────────────────────────────────────────────────────
STREAM_ENABLED = os.getenv("STREAM_ENABLED", "true").lower() == "true"

# ── Langfuse tracing (self-hosted) ─────────────────────────────────────────────
LANGFUSE_ENABLED = os.getenv("LANGFUSE_ENABLED", "false").lower() == "true"
LANGFUSE_PUBLIC_KEY = os.getenv("LANGFUSE_PUBLIC_KEY", "")
LANGFUSE_SECRET_KEY = os.getenv("LANGFUSE_SECRET_KEY", "")
LANGFUSE_HOST = os.getenv("LANGFUSE_HOST", "http://localhost:3000")
LANGFUSE_PROMPT_LIMIT = int(os.getenv("LANGFUSE_PROMPT_LIMIT", "4000"))  # max chars in trace input/output

	"""
	Central configuration for the Agentic RAG system.

	Edit AGENT_* env vars to tune behavior without touching code.
	All prompts live in prompts.py (single source of truth).
	"""

	import os

	# ── Agent loop ──────────────────────────────────────────────────────────────────
	AGENT_MAX_RETRIES = int(os.getenv("AGENT_MAX_RETRIES", "2"))
	AGENT_SEARCH_TIMEOUT = float(os.getenv("AGENT_SEARCH_TIMEOUT", "15.0"))
	AGENT_GENERATION_TIMEOUT = float(os.getenv("AGENT_GENERATION_TIMEOUT", "180.0"))
	AGENT_MAX_CONTEXT_TOKENS = int(os.getenv("AGENT_MAX_CONTEXT_TOKENS", "4000"))
	AGENT_DEFAULT_K = int(os.getenv("AGENT_DEFAULT_K", "10"))

	# ── Mem0 (self-hosted, OSS) ───────────────────────────────────────────────────
	MEM0_ENABLED = os.getenv("MEM0_ENABLED", "true").lower() == "true"
	MEM0_USER_ID_DEFAULT = os.getenv("MEM0_USER_ID_DEFAULT", "default_session")
	MEM0_MAX_MEMORIES = int(os.getenv("MEM0_MAX_MEMORIES", "5"))
	MEM0_COLLECTION_NAME = os.getenv("MEM0_COLLECTION_NAME", "mem0_cve_memories")

	# ── Web search (optional, requires Tavily API key) ─────────────────────────────
	WEB_SEARCH_ENABLED = os.getenv("WEB_SEARCH_ENABLED", "false").lower() == "true"

	# ── Grading backend ────────────────────────────────────────────────────────────
	# "reranker" — use Jina Reranker v3 (fast, no tokens consumed, recommended)
	# "llm" — use batched LLM call (slower, costs tokens, fallback)
	GRADE_BACKEND = os.getenv("GRADE_BACKEND", "reranker")

	# ── Reranker (Jina Reranker v3) ────────────────────────────────────────────────
	RERANKER_BACKEND = os.getenv("RERANKER_BACKEND", "local") # "local" \| "api"
	RERANKER_MODEL = os.getenv("RERANKER_MODEL", "jinaai/jina-reranker-v3")
	RERANKER_THRESHOLD = float(os.getenv("RERANKER_THRESHOLD", "0.5")) # sigmoid-normalized; 0.5 = neutral

	# ── Grading thresholds (LLM path) ──────────────────────────────────────────────
	GRADE_RELEVANCE_THRESHOLD = float(os.getenv("GRADE_RELEVANCE_THRESHOLD", "1")) # min relevant docs
	GRADE_MIN_PASSING_SCORE = float(os.getenv("GRADE_MIN_PASSING_SCORE", "0.5"))

	# ── Streaming ──────────────────────────────────────────────────────────────────
	STREAM_ENABLED = os.getenv("STREAM_ENABLED", "true").lower() == "true"

	# ── Langfuse tracing (self-hosted) ─────────────────────────────────────────────
	LANGFUSE_ENABLED = os.getenv("LANGFUSE_ENABLED", "false").lower() == "true"
	LANGFUSE_PUBLIC_KEY = os.getenv("LANGFUSE_PUBLIC_KEY", "")
	LANGFUSE_SECRET_KEY = os.getenv("LANGFUSE_SECRET_KEY", "")
	LANGFUSE_HOST = os.getenv("LANGFUSE_HOST", "http://localhost:3000")
	LANGFUSE_PROMPT_LIMIT = int(os.getenv("LANGFUSE_PROMPT_LIMIT", "4000")) # max chars in trace input/output