llama-index-callbacks-hdp 0.1.1

HDP (Human Delegation Provenance) — cryptographic authorization provenance for LlamaIndex agents

llama-index-callbacks-hdp

HDP (Human Delegation Provenance) integration for LlamaIndex — cryptographic authorization provenance for agents and RAG pipelines.

HDP answers the question that observability tools like Arize Phoenix and Langfuse cannot: who authorized this agent run, under what scope, and can you prove it offline?

Every tool call, retrieval step, and LLM invocation is recorded in a tamper-evident, cryptographically signed delegation chain. The chain is fully verifiable offline — no network calls, no central registry.

Installation

pip install llama-index-callbacks-hdp

Usage

Option 1 — Modern instrumentation dispatcher (LlamaIndex ≥0.10.20)

from llama_index.callbacks.hdp import HdpInstrumentationHandler, HdpPrincipal, ScopePolicy

HdpInstrumentationHandler.init(
    signing_key=ed25519_private_key_bytes,
    principal=HdpPrincipal(id="[e-mail]", id_type="email"),
    scope=ScopePolicy(
        intent="Research pipeline",
        authorized_tools=["web_search", "retriever"],
        max_hops=10,
    ),
    on_token_ready=lambda token: print(token["header"]["token_id"]),
)

Option 2 — Legacy CallbackManager

from llama_index.callbacks.hdp import HdpCallbackHandler, HdpPrincipal, ScopePolicy
from llama_index.core import Settings
from llama_index.core.callbacks import CallbackManager

handler = HdpCallbackHandler(
    signing_key=ed25519_private_key_bytes,
    principal=HdpPrincipal(id="[e-mail]", id_type="email"),
    scope=ScopePolicy(intent="Research pipeline"),
)
Settings.callback_manager = CallbackManager([handler])

Option 3 — Node postprocessor (inline retrieval enforcement)

from llama_index.callbacks.hdp import HdpNodePostprocessor

postprocessor = HdpNodePostprocessor(
    signing_key=ed25519_private_key_bytes,
    strict=False,
    check_data_classification=True,
)
query_engine = index.as_query_engine(node_postprocessors=[postprocessor])

Verifying a token

from llama_index.callbacks.hdp import verify_chain

result = verify_chain(token_dict, public_key_bytes)
if result.valid:
    print(f"Chain verified: {result.hop_count} hops")

What makes HDP different from Arize/Langfuse?

Capability	Arize / Langfuse	HDP
Records what happened	✓	✓
Records who authorized it	✗	✓
Cryptographically signed	✗	✓
Verifiable offline	✗	✓
Scope enforcement	✗	✓
No central registry	n/a	✓

License

Apache-2.0