Retrieval-augmented generation (RAG) is the workhorse pattern of enterprise AI: index your documents, retrieve the relevant passages, let a language model answer with citations. The reference architectures, however, assume infrastructure that Nigerian deployments cannot take for granted — stable broadband, low-latency cloud access, and uninterrupted power. A RAG assistant that works beautifully in a Lagos head office demo can become unusable in a branch on a congested link. These are the patterns we apply to keep systems working when infrastructure does not cooperate.
Design for the p95 network, not the demo network
The single most important discipline is measuring and designing for the 95th-percentile network conditions of your actual users. A round trip to an offshore model API that takes 800ms on fibre can take 8 seconds on a saturated branch VSAT link — and a chat interface that streams tokens over that link compounds the pain. Practical consequences: keep payloads small (retrieve passages, not documents; compress aggressively), stream responses so users see progress within the first second, and set timeout-and-retry budgets that reflect real conditions rather than defaults.
Hybrid retrieval beats embedding-only retrieval on constrained links
Pure vector search requires an embedding call per query — another network round trip if your embedding model is remote. A hybrid design pairs BM25 keyword retrieval, which runs locally and instantly, with vector search as an enhancer. When the network is healthy, hybrid scores fuse both signals; when it degrades, the system falls back to keyword-only retrieval and still returns grounded answers. Users experience slightly less semantic finesse rather than an error message. The same logic argues for hosting the embedding model in-country or on-premise where data-residency rules already push you that way.
Cache at every layer, and treat the cache as a feature
In enterprise knowledge work, query distributions are heavily repetitive — policy questions, product terms, procedural how-tos. A semantic cache that recognises paraphrases of previously answered questions can serve a large share of traffic without touching the model at all: instant answers, zero bandwidth, zero inference cost. Layer caches deliberately: an exact-match cache for repeated queries, a semantic cache above it, and a passage cache so popular documents are served from memory. For field deployments, pre-warming caches with the fifty most common questions turns the worst-connectivity case into the best-experience case.
Graceful degradation is an architecture, not an apology
Define explicit service tiers and make the system state visible. Tier one: full RAG with generation and citations. Tier two: retrieval-only — the system returns the most relevant passages with highlighting, no generated synthesis. Tier three: cached answers only, clearly labelled with their freshness. A status indicator that says "operating in offline mode — showing source passages" preserves trust in a way that a spinning loader never will. Crucially, tiers two and three still deliver real value: in document-heavy workflows, fast passage retrieval is often 80% of the benefit.
Power reality shapes deployment topology
Bandwidth is only half the constraint; the other half is power. An on-premise inference server is only as available as the generator schedule behind it. This pushes toward asymmetric designs: a lightweight, resilient edge (retrieval, cache, UI) that runs on modest hardware with UPS backing, and heavier generation capacity in a well-powered data centre or sovereign cloud region, reached opportunistically. Where the use case demands full offline capability, small quantised models running on-CPU now handle summarisation and extraction credibly — a pattern we use in our document intelligence deployments for sites where connectivity is a privilege, not a guarantee.
The uncomfortable truth: infrastructure constraints improve the engineering
Every pattern above — caching, hybrid retrieval, tiered degradation, edge-weighted topology — also cuts inference costs and improves latency in perfect conditions. Designing for Nigerian infrastructure does not produce a lesser system; it produces a more disciplined one. Our engineering accelerators embed these patterns as defaults, which is why a Muller Global RAG deployment is scoped in weeks rather than quarters: the hard lessons are already in the reference architecture.
About Muller Global
Muller Global is an AI Advisory & Engineering firm based in Abuja, taking Nigerian enterprises from AI ambition to working systems through consulting, advisory retainers, training, deployment, and engineering.
The Muller Weekly Briefing
One email a week on AI for Nigerian enterprises — what matters, minus the hype.