Technical Governance & Systems Design Proposals

Engineering RFC Proposal Repository

Formal architectural Request for Comments (RFC) specifications documenting technical motivation, trade-offs, security controls, and canary deployment plans.

RFC Engineering Governance & Architecture Proposals

Engineering RFC Proposal & Governance Studio

Production-grade engineering RFCs detailing technical motivation, rejected alternatives, security implications, and zero-downtime canary rollout plans.

Staff Engineering Standards

RFC-001Approved & Implemented

Subsurface Geological Vector Ingestion & PostGIS Cosine Indexing

Author: Abdlrrahman Shibani • Date: 2025-11-14

1. Executive Summary:

Defines the architecture for asynchronous chunked LAS well-log ingestion, transformer embeddings generation (GeoBERT), and combined PostGIS spatial + pgvector similarity indexing.

2. Problem Statement & Motivation:

Legacy oil exploration platforms suffered from 12+ second query latencies across North Sea geological surveys. Unifying PostGIS spatial bounding boxes with 768-dim vector cosine similarity cut query times to <40ms.

3. Proposed Architectural Solution:

1. Fast binary LAS file parser using NumPy memory mapping. 2. GeoBERT transformer inference pipeline with ONNX Runtime CPU fallback on GPU OOM. 3. HNSW cosine index inside PostgreSQL 16 using pgvector extension.

4. Rejected Alternatives & Trade-Off Analysis:

❌Elasticsearch dense_vector: Rejected due to memory overhead and lack of native PostGIS spatial topology operations.

❌Milvus standalone cluster: Rejected to avoid multi-database operational complexity and transactional consistency overhead.

5. Security & Isolation Controls:

Tenant-scoped vector namespaces enforced via PostgreSQL Row-Level Security (RLS) to prevent cross-operator subsurface lease data leakage.

6. Phased Rollout & Canary Verification:

Phase 1: Dual-write shadow indexing in staging. Phase 2: 10% canary traffic verification against 50k well logs. Phase 3: Full production promotion with 0 downtime.