palveron-langchain

PALVERON AI Governance adapter for LangChain — automatic policy checks, PII masking, and audit trails for every LLM call in your pipeline.

---

Add one callback handler. Every prompt, every tool call, every LLM output gets checked against your PALVERON governance policies. PII is detected. Secrets are caught. Blocked requests raise exceptions before they reach the model.

Installation

pip install palveron-langchain

Quick Start

from palveron_langchain import PalveronCallbackHandler
from langchain_openai import ChatOpenAI

# Create the governance handler
handler = PalveronCallbackHandler(api_key="pv_live_xxx")

# Attach to any LangChain component
llm = ChatOpenAI(model="gpt-4o", callbacks=[handler])

# Every call is now governed
result = llm.invoke("Transfer $50,000 to account DE89370400440532013000")
# → PalveronGovernanceError: Blocked — PII detected (IBAN)

What Gets Checked

Event	When	What
on_llm_start	Before text completion	Prompt text
on_chat_model_start	Before chat model call	All messages concatenated
on_tool_start	Before tool execution	Tool name + input
on_llm_end	After generation (optional)	LLM output text

Configuration

handler = PalveronCallbackHandler(
    api_key="pv_live_xxx",
    base_url="https://gateway.palveron.com",  # On-prem endpoint
    check_prompts=True,       # Check inputs before LLM (default: True)
    check_outputs=False,      # Check LLM outputs (default: False)
    check_tools=True,         # Check tool inputs (default: True)
    raise_on_block=True,      # Raise exception on BLOCKED (default: True)
    fail_open=False,          # Allow calls when gateway down (default: False)
    metadata={"team": "ml"},  # Extra metadata on every trace
)

Behavior on Decisions

Decision	raise_on_block=True	raise_on_block=False
ALLOWED	Call proceeds	Call proceeds
MODIFIED	Call proceeds, PII redaction logged	Call proceeds, PII redaction logged
BLOCKED	Raises PalveronGovernanceError	Logs warning, call proceeds
ERROR	Depends on fail_open	Depends on fail_open

Note: LangChain callbacks are observational — they cannot modify prompts in-flight. For full input rewriting (PII masking before the LLM sees it), use the PALVERON Gateway Proxy.

Governance Records

Access the audit trail programmatically:

# After running your chain
print(f"Blocked: {handler.blocked_count}")
print(f"Trace IDs: {handler.trace_ids}")

for record in handler.records:
    print(f"{record.event}: {record.decision} ({record.latency_ms:.0f}ms) — {record.trace_id}")

Error Handling

from palveron_langchain import PalveronCallbackHandler, PalveronGovernanceError

handler = PalveronCallbackHandler(api_key="pv_live_xxx")

try:
    result = llm.invoke("Send SSN 123-45-6789 to the client", config={"callbacks": [handler]})
except PalveronGovernanceError as e:
    print(e.decision)   # "BLOCKED"
    print(e.trace_id)   # "trc_abc123"
    print(e.reason)     # "PII detected: Social Security Number"

With Agents

from langchain.agents import create_react_agent

handler = PalveronCallbackHandler(api_key="pv_live_xxx", check_tools=True)

agent = create_react_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, callbacks=[handler])

# Both LLM calls AND tool executions are governed
result = agent_executor.invoke({"input": "Delete all customer records"})

With Chains (LCEL)

from langchain_core.prompts import ChatPromptTemplate

handler = PalveronCallbackHandler(api_key="pv_live_xxx")

chain = ChatPromptTemplate.from_template("Summarize: {text}") | llm
result = chain.invoke({"text": sensitive_doc}, config={"callbacks": [handler]})

Requirements

• Python 3.9+
• palveron-sdk >= 0.5.0
• langchain-core >= 0.2.0

Links

• Documentation
• PALVERON Dashboard
• palveron-sdk (Python)
• GitHub

License

Apache 2.0