[ DevConf 2026 ] RAG that works in South Africa: cheap, fast, and compliant approaches any company can implement 29-Jun-26 By Abed Matini 1 Senior Backend Developer at 

[ DevConf 2026 ] 29-Jun-26 2 How Retrieval Augmented Generation (RAG) Works? 

[ DevConf 2026 ] Why global RAG advice breaks in South Africa? • Regulatory (POPIA) Strict rules restrict sending sensitive data to foreign cloud APIs. • Currency Risk ZAR/USD volatility makes cloud token costs unpredictable. • Connectivity Load shedding and unreliable internet disrupt cloud-only operations. • Legacy Infrastructure 29-Jun-26 Scanned PDFs and messy tables break standard parsers. 3 

[ DevConf 2026 ] The South African reality just became draft law • Draft SA National AI Policy gazetted on 10 April 2026. • Public comments close: 10 June 2026. • The 3-Year Plan: ✓ Year 1 (2025/26): Finalize Policy & target unacceptable risks. ✓ Year 2 (2026/27): Guidelines & high-risk regulations. 29-Jun-26 ✓ Year 3 (2027/28): Full implementation & audits. 4 

[ DevConf 2026 ] Draft AI Policy Framework for RAG • Data Sovereignty (9.3.2) Local RAG is the primary defense against prohibited RAG citations provide the "interpretable and "data extraction". contestable" architecture required for high-risk AI. 29-Jun-26 • Human-in-the-loop (9.6.1) 5 • Explainability (9.6.2) • Liability (7.1) Agentic frameworks must include manual Hallucinations are now legal liabilities, overseen by checkpoints for critical decisions. the new AI Ombudsperson. 

[ DevConf 2026 ] The POPIA & connectivity matrix Local Deployment Required for highly sensitive data under strict POPIA rules or areas with poor internet. Cloud Deployment Safe for low-risk or public data backed by reliable fibre connectivity. Hybrid RAG (Default) 29-Jun-26 Routes safe queries to the cloud while strictly keeping sensitive data local. 6 

[ DevConf 2026 ] The Reality of ZAR vs. USD (Cost Over Time) • Cloud RAG Low initial setup, but long-term scaling costs are inflated by ZAR/USD exchange rate volatility. • Local RAG Higher upfront hardware investment, but guarantees stable and predictable operational costs over time. • Hybrid Advantage: 29-Jun-26 Balances the initial hardware investment with controlled, predictable monthly costs. 7 

[ DevConf 2026 ] 29-Jun-26 8 The RAG Maturity Roadmap 

[ DevConf 2026 ] Phase 1: Secure Local RAG The Concept: A 100% private, offline search engine built on RBAC. The Trigger: Zero API budget + strict POPIA mandates. The Governance Trap: Local storage is not enough; you must protect against internal threats. The Fix (Locks on the inside): • Never raw-dump data; access control precedes retrieval. • Attach metadata tags (e.g., Dept: HR, Clearance: L3) to vector chunks. { "chunk_id": "doc1_part3", 29-Jun-26 "text": "...", 9 "allowed_groups": ["hr","managers"], "owner": "john.doe“ } • Verify Active Directory credentials first, making unauthorized data mathematically invisible. Suggested Stack: Llama 3 (8B) via Ollama + Qdrant/Milvus 

[ DevConf 2026 ] Phase 2: Solving the Connectivity & Budget Squeeze (Chunk-Rank) • The Concept: A "smart filter" that mathematically scores and filters relevant documents before sending them to the LLM. • The Trigger: Basic search retrieves irrelevant data, wasting expensive compute resources. • The Financial Win: By actively preventing the system from processing useless data, you avoid massive token usage and mitigate ZAR/USD exchange rate volatility. Suggested Stack: • Combine BM25 Keyword Search + Vector Search. 29-Jun-26 • Local Reranker like BAAI/bge-reranker-v2-m3. 10 

[ DevConf 2026 ] Phase 3: Solving the "Ugly Data" Crisis (Multi-Vector RAG) • The Concept: Search easy-to-read text summaries, but retrieve the original "messy" visual files for the AI to analyze. • The Trigger: Standard parsers break when reading scanned Home Affairs IDs and blurry SARS ITA34 tax documents. 29-Jun-26 • How it Works: The system creates a clean text summary of the scanned PDF to make it searchable, but hands the raw image to a Vision LLM to preserve rich visual context. 11 • Suggested Stack: Unstructured.io / ColPali for parsing, Llama-3.2-Vision (Local). 

[ DevConf 2026 ] Phase 4: Solving Complex Cross-Referencing (Multi-Step RAG) • The Concept: Breaking complex questions into smaller "search missions". • The Trigger: The AI fails when asked to compare disparate data, like "Compare 2023 vs 2024 Policy". • How it Works: • Step 0: "Query Rewriting” • Finance: Splits "Q1 vs Q2" into two distinct retrievals. 29-Jun-26 • HR: Compares "2023 vs 2024 Policy" sequentially. 12 • Suggested Stack: LlamaIndex (Query Pipelines) or LangChain. 

[ DevConf 2026 ] 29-Jun-26 Advanced Insight and Action 13 

[ DevConf 2026 ] Phase 5: Solving the Macro-Summary Bottleneck (Map-Reduce RAG) • The Concept: Summarizing massive document batches simultaneously without hitting token limits. • The Trigger: Stakeholders need "big picture" macro-level insights across 500+ reports. Process: • Map: Batch documents -> Generate mini-summaries. • Reduce: Synthesize mini-summaries -> Final master report. 29-Jun-26 • Suggested Stack: LangChain MapReduceDocumentsChain 14 

[ DevConf 2026 ] Phase 6: Solving Hidden Compliance Webs (GraphRAG) • The Concept: Building a "detective string-board" Knowledge Graph to uncover hidden structural relationships. • The Trigger: Mapping complex webs like procurement fraud, vendor relationships, or POPIA impact zones. How it Works: • Extraction: LLMs extract Entities (People, Systems) and Relationships (Ownership, Access). • Knowledge Graph: Connects the dots across disparate, disconnected documents. 29-Jun-26 • Suggested Stack: Microsoft GraphRAG, Neo4j, LlamaIndex Property Graphs , or PostgreSQL + Apache AGE for standard Cypher queries. 15 Warning: LLM-generated GraphRAG is incredibly expensive, slow, difficult to maintain, and often yields weak production ROI. 

[ DevConf 2026 ] Phase 7: Solving the Automation Bottleneck (Agentic RAG) • The Concept: Giving the AI "hands" to interact with company systems (e.g., executing SQL, logging Jira tickets, emailing site managers). 29-Jun-26 • Suggested Stack: Gemini 2.5 Pro or GPT-4o with LangChain/CrewAI. 16 • The Warning: Because we established metadata filtering and RBAC in Phase 1, the Agent cannot go rogue on sensitive systems. Human-in-the-loop checkpoints are mandatory. 

[ DevConf 2026 ] Phase 7b: The Compliance Key for Agents (Agentic Consent) Ensuring agents act with us, not instead of us. • Dynamic Consent : Move from static checkboxes to context-aware, living contracts. • Identity Verification: Cryptographically verify agent actions against Active Directory clearance. • Granular & Transaction-Based Access: Restrict agents to specific actions (e.g., read-only) for a 29-Jun-26 narrow transaction window. 17 • Just-in-Time Prompting (The Human-in-the-Loop): Mandate explicit human consent before processing sensitive data. • POPIA alignment: Ensure agent actions are transparent, revokable, and customizable by the user. 

SAMPLE FOOTER TEXT We locked our data in a local vault to comply with POPIA. 29-Jun-26 But who is protecting your data from the inside? 18 

[ DevConf 2026 ] The governance trap (context before retrieval) The Risk: AI bypassing standard folder restrictions to answer unauthorized internal queries. The Fix: Metadata + RBAC (Role-Based Access Control) • Attach permission tags (Department, Clearance Level) to stored vector chunks. • Mandated by AI Policy: Data Minimisation & Purpose Limitation The Golden Rule: Access control must precede retrieval; never 29-Jun-26 raw-dump data. 19 Suggested Stack: Qdrant, Milvus, or Weaviate 

[ DevConf 2026 ] 29-Jun-26 20 Validation and Practical Next Steps 

[ DevConf 2026 ] Proving It Works (Benchmarking) The Concept: Mathematically proving your AI is accurate, safe, and hallucination-free. The Trigger: The executive team asks, "How do we know it works?" • Step 1: The Golden Dataset • 30–50 difficult, real-world business questions with expert-verified answers. • Step 2: The RAG Triad 1. Context Relevance: Did it retrieve the right document? 2. Faithfulness: Did it stick strictly to the facts (no hallucinations)? 3. Answer Relevance: Did it actually answer the user's question? 29-Jun-26 Cost hack: Use Regex for citation checking. 21 Suggested Stack: Ragas or TruLens. Mandated by AI Policy: Achieving 'Sufficient Explainability' 

[ DevConf 2026 ] The 5-Minute Sanity Check (No Code Required) 29-Jun-26 The Goal: Troubleshooting your Local RAG database. 22 • Test 1 (Blind Test): Ask the question without data. It should fail. (Ensures the AI isn't just guessing from its training). • Test 2 (Copy-Paste Test): Paste the entire document into the chat. If it passes, the LLM is smart enough to understand your data • Test 3 (RAG Test): Run it through the system. If it fails now (but passed Test 2), your chunking/retrieval is the problem. 

[ DevConf 2026 ] The South African RAG Cheat Sheet (Summary) 1. Simple RAG: Zero Budget / POPIA? Run it locally. 2. Chunk-Rank: Accuracy failing? Add a local Reranker & Hybrid Search. 3. Multi-Vector: Scans & Tables? Summarize for search, read raw visuals. 4. Multi-Step: Need Comparisons? Split prompts into missions. 5. Map-Reduce: 500+ Document Summary? Batch and merge summaries. 6. GraphRAG: 29-Jun-26 7. Agentic RAG: 23 Hidden Fraud/Links? Build a Relationship Graph. Need to Automate? Connect tools (SQL, Email, Jira). 

[ DevConf 2026 ] Monday morning action plan Step 1: The POPIA & Data Audit Assess sensitivity to rule out pure Cloud RAG. Step 2: The "Ugly Data" Test Run text extraction tests on messy, real-world docs (IDs, SARS ITA34). Step 3: The Local Sandbox Deploy Llama 3 locally, ensuring metadata tagging is active for compliance. (Chunking rule: Use 512 tokens, 25% overlap) Step 4: The Executive Pitch Run a "Golden Dataset" to mathematically prove zero hallucinations and 29-Jun-26 calculate OPEX savings against volatile USD cloud pricing. 24 Prove Safety: Use TruLens to generate a printable report proving Faithfulness— mathematically. 

Thank you. https://www.linkedin.com/in/matini 

Scan to give Feedback https://www.devconf.co.za/rate-capetown