graph-abstract 0.1.0

A wrapper for LangGraph agents that handles observability, memory, and guardrails

graph-wrap

A local-first Python library that provides zero-config PostgreSQL checkpointing and telemetry logging for LangGraph agents simply by replacing your standard StateGraph import.

---

Installation

Core Observability and Checkpointing

To install the core package with database checkpointing and telemetry logging:

pip install graph-wrap

With Optional UI Console

To install the package along with the Streamlit-based visualization console:

pip install "graph-wrap[ui]"

---

Quick Start

Replace your standard LangGraph StateGraph import with graph_wrap:

from typing import Dict, Any
from typing_extensions import TypedDict
from graph_wrap import StateGraph

class AgentState(TypedDict):
    messages: list[str]

def my_node(state: AgentState) -> Dict[str, Any]:
    return {"messages": ["hello"]}

db_uri = "postgresql://postgres:postgres@localhost:5432/my_database"
graph = StateGraph(AgentState, db_uri=db_uri)
graph.add_node("my_node", my_node)
graph.set_entry_point("my_node")
graph.set_finish_point("my_node")

compiled = graph.compile()

When you call compiled.ainvoke, the library dynamically setups Postgres checkpoint tables and logs execution traces (chains, tools, and LLMs) into the database under the configured thread_id.

---

Observability UI Console

If you installed graph-wrap with the ui extra, you can launch the console to inspect threads, visualize trace events, and view state checkpoints.

Run the console via python:

python -m graph_wrap --db-uri postgresql://postgres:postgres@localhost:5432/my_database

Or run the command line executable:

graph-wrap-ui --db-uri postgresql://postgres:postgres@localhost:5432/my_database

---

Custom Guardrails

graph-wrap provides custom inbound and outbound safety guardrails. Evaluations for prompt injection, content safety, and hallucination are performed using LangChain's native structured output bindings for Ollama and OpenAI. Local regex-based PII redaction runs locally with zero LLM costs.

Quick Example

from typing_extensions import TypedDict
from graph_wrap import StateGraph, GuardrailConfig, GuardrailProvider

class AgentState(TypedDict):
    messages: list[str]
    context: str

safety_config = GuardrailConfig(
    inbound_provider=GuardrailProvider.OLLAMA,
    safety_model="llama3.1:8b",
    check_prompt_injection=True,
    
    outbound_provider=GuardrailProvider.OLLAMA,
    eval_model="llama3.1:8b",
    check_hallucination=True,
    
    redact_pii=True,
    fallback_message="I cannot answer that, please try to rephrase your question."
)

workflow = StateGraph(
    AgentState, 
    db_uri="postgresql://postgres:postgres@localhost:5432/my_database",
    guardrails=safety_config
)

Node-Specific Guardrails

To apply guardrails to specific nodes only, pass a dictionary mapping node names to GuardrailConfig objects:

workflow = StateGraph(
    AgentState,
    db_uri="postgresql://postgres:postgres@localhost:5432/my_database",
    guardrails={
        "Trader": GuardrailConfig(
            outbound_provider=GuardrailProvider.OLLAMA,
            eval_model="llama3.1:8b",
            check_hallucination=True
        )
    }
)

Config Options

Parameter	Type	Default	Description
inbound_provider	GuardrailProvider	None	Provider for inbound validation (OLLAMA or OPENAI).
safety_model	str	None	Model name to use for inbound safety/prompt injection checks.
check_prompt_injection	bool	False	Enable inbound prompt injection and jailbreak detection.
outbound_provider	GuardrailProvider	None	Provider for outbound validation (OLLAMA or OPENAI).
eval_model	str	None	Model name to use for outbound hallucination or safety checks.
check_hallucination	bool	False	Enable outbound hallucination check comparing node outputs to graph context.
redact_pii	bool	False	Enable local regex-based redaction of emails, credit cards, SSNs, and phone numbers.
fallback_message	Optional[str]	"I cannot answer..."	The message returned when a guardrail is tripped. Set to None to raise a GuardrailValidationError instead.