voyage-ledgerproof 0.1.1

LedgerProof adapter for the Voyage AI Python SDK. Side-channel transparency receipts for EU AI Act Article 50 RAG pipeline evidence.

ledgerproof-voyage

LedgerProof adapter for the Voyage AI Python SDK (embeddings + rerank).

Emits side-channel cryptographic transparency receipts that capture the upstream retrieval evidence (embeddings, rerank scores) feeding a downstream chatbot, suitable as an evidence layer for the EU AI Act Article 50 RAG pipeline scenario.

This adapter is not endorsed by Voyage AI or MongoDB. It wraps the official voyageai SDK; it is not affiliated with Voyage AI Inc. or MongoDB Inc.

Why an embeddings adapter for an Article 50 protocol?

Voyage AI is upstream RAG infrastructure. Calls to voyageai.Client.embed() and voyageai.Client.rerank() are not themselves direct Article 50(1) "AI system intended to interact directly with natural persons" surfaces — they are retrieval-side substrate.

But they are the evidence trail for any RAG pipeline whose downstream chatbot output is under Article 50. When a regulator (or an internal auditor, or a litigant) asks the load-bearing question "what sources did the model see, and how did they get there?", the answer is a chain:

user query
  -> Voyage embed(query)              [embedding_inference/v1]
  -> Voyage embed(documents)          [embedding_inference/v1]
  -> ANN retrieval (vector DB)        [out of scope]
  -> Voyage rerank(query, candidates) [rerank_inference/v1]
  -> downstream chatbot turn          [bound by rag_pipeline_evidence/v1]

This adapter signs each link of that chain locally (constraint C4) so the full provenance is verifiable offline from the stored signed receipts. The downstream chatbot turn is captured by the matching LedgerProof chat adapter (ledgerproof-openai, ledgerproof-anthropic, ledgerproof-cohere, etc.) and both are stitched together by a rag_pipeline_evidence/v1 envelope.

This positioning matters because most Tier-1 EU enterprise AI deployments under Article 50 enforcement (2 August 2026) are RAG-shaped, and the regulator's first incident-response question is invariably "what evidence does the model have?". A signed Voyage embedding + rerank receipt is the cryptographic answer.

5-minute quickstart

pip install ledgerproof-voyage
export VOYAGE_API_KEY=...

from voyage_ledgerproof import LedgerProofVoyage, LogEmitter

client = LedgerProofVoyage(
    deployer_id="acme-corp-eu",
    emitter=LogEmitter(),
)

result = client.embed(
    texts=["Article 50 transparency obligations apply from 2 August 2026."],
    model="voyage-3-large",
    input_type="document",
)

print(result.embeddings[0][:4], "...")
# Receipt has already been emitted to the side-channel.

The Voyage EmbeddingsObject is returned unchanged. The receipt is emitted to the side channel only (C7).

Three schemas

Schema	Article 50 mapping	What it binds
embedding_inference/v1	Supporting infrastructure	per-input SHA-256, per-vector SHA-256, vector dim, model_id, input_type
rerank_inference/v1	Supporting infrastructure	query SHA-256, per-candidate SHA-256, relevance scores, post-rerank order, model_id
rag_pipeline_evidence/v1	Article 50(1) variant	upstream embed + rerank receipt hashes -> downstream chat receipt hash + user query hash

embedding_inference/v1 and rerank_inference/v1 carry regulatory_context.article_50_paragraph = "supporting" — they are not Article 50(1) artefacts in themselves. rag_pipeline_evidence/v1 carries "1" because that is the regulated user-facing turn.

Three integration patterns

1. Client wrapper (recommended) — LedgerProofVoyage wraps voyageai.Client and intercepts embed() and rerank(). Sync and async (LedgerProofAsyncVoyage) supported.
2. Decorator — @lpr_track_embed(deployer_id=..., model=...) and @lpr_track_rerank(deployer_id=..., model=...) for user-defined functions wrapping Voyage calls.
3. Manual emission — emit_embedding_receipt(...), emit_rerank_receipt(...), and emit_rag_pipeline_receipt(...) for full control inside a custom RAG orchestration layer.

See examples/:

• examples/01_voyage_embed_quickstart.py — vanilla embed()
• examples/02_voyage_rerank.py — rerank() with relevance-score binding
• examples/03_rag_pipeline_evidence.py — full embed -> rerank -> (downstream chat stub) -> rag_pipeline_evidence/v1

Architectural discipline (C1–C8)

This adapter is implemented under the LedgerProof protocol's load-bearing constraints:

• C1: No claim of regulator endorsement. No claim of Article 40 presumption of conformity. Not endorsed by Voyage AI Inc. or MongoDB Inc. This adapter wraps the official Voyage Python SDK; it is not affiliated with either company.
• C4: Local verification only. The adapter does not phone home to LedgerProof servers during normal operation. Signed receipts verify offline given the public key.
• C6: Stream-aware signing primitives are exposed for downstream chat adapters. Voyage's own surfaces (embed, rerank) are non-streaming today.
• C7: Side-channel emission only. The adapter cannot and does not modify the Voyage response payload. The EmbeddingsObject / RerankingObject is returned unchanged.

GDPR posture

Receipts reference content by SHA-256 hash only. The adapter does not store:

• raw input texts to embed()
• raw document text fed to rerank()
• raw embedding vectors (the vector is canonicalized to big-endian float64 and hashed; the hash is stored, not the vector)
• raw user queries fed into the downstream chatbot

Identifiers (deployer_id, document_id, tool_call_id) are length-bounded and character-set- restricted (no free-form PII). The deployer is responsible for ensuring the inputs they feed in (e.g. document_id values) do not themselves leak PII.

Vector canonicalization

Embedding vectors are canonicalized to IEEE-754 big-endian float64 byte strings before hashing. This is portable across machines, Python versions, and operating systems. A verifier with the original vector can recompute SHA-256(canonicalize_vector(vector)) and compare to vector_sha256_hex in the receipt.

Voyage exposes output_dtype= for higher-precision dtypes (int8, uint8, binary, ubinary); the adapter records the requested output_dtype on VoyageModelRef.output_dtype for transparency, but always canonicalizes through float64 for one stable wire format.

Scope disclaimer

LedgerProof provides an evidence layer for Article 50 transparency obligations only.

It does not cover:

• Article 9 (risk management system)
• Article 10 (data and data governance)
• Article 13 (transparency and information to deployers)
• Article 15 (accuracy, robustness, and cybersecurity)
• Article 72 (post-market monitoring)

LedgerProof does not confer presumption of conformity under Article 40. LedgerProof is not endorsed by the European Commission, the AI Office, any national competent authority, Voyage AI Inc., or MongoDB Inc.

This adapter wraps the official voyageai Python SDK; it is not affiliated with Voyage AI Inc. or MongoDB Inc.

License

Apache 2.0. Copyright 2026 LedgerProof Foundation (in formation: US 501(c)(3) Delaware + Dutch Stichting EU subsidiary).