Welcome

Not here to sound smart.
Here to make things work.

AI・Tech・Business・Web3・Travel

See all posts
Scroll

Latest

View all
AI Jun 12, 2026

From RAG to Enterprise RAG Part 11 | After Finishing a Round: 3 Things I Took Away From This Project

Parts 02-10 walked through the complete journey from minimum viable to production. This series closer articulates 3 things I only understood after doing them: (1) collapsing 11 capability flags into 5 query modes—the shift from feature-oriented to system-oriented design, (2) re-encoding the entire learning path after one round—the teacher's own re-design, (3) what I chose NOT to build—the deferred list is more honest than the done list. Full series closure. Part 01 is the interactive demo entry; Part 11 is the coordinate system.

Read more EN
AI Jun 12, 2026

From RAG to Enterprise RAG Part 10 | Deploying My RAG: 10 Pits I Fell Into

Going from a local prototype to a public API isn't the final step—it's another exam. This article walks through 11 real deployment pitfalls (Oracle VM capacity, Docker env shadowing, Compose healthcheck cascading failures, Nginx + Cloudflare HTTPS trust chain, Qdrant payload index bootstrap, etc.) across four infrastructure layers, with prevention frameworks for each. Personal architecture choices (Oracle VM / Qdrant Cloud / Cloudflare / Docker Compose) explicitly marked as 'my choices, not general advice'.

Read more EN
AI Jun 11, 2026

From RAG to Enterprise-Grade RAG Part 09 | Make documents actually enter, stay governed, and stay queryable: from ingestion to document APIs

Part 08 bundled four capabilities into five query modes, but no matter how strict the query-side routing gets, a broken document side still breaks the system — user A pulling user B's contract, header spoofing to impersonate other users, deleting a file and losing the chunks. This post walks through the seven stations a document actually travels after upload (parsing → raw storage → Postgres metadata → ingestion queue → auth/permission → document APIs → citation viewer), with the focus on the often-overlooked claim that the Vector DB cannot be the source of truth, four ACL failure scenarios (header spoofing / no tenant filter / app-layer filter with no RLS / JWT without membership check), and the citation viewer's payload as the last mile of explainability. Document-side seven stations (Part 09) + query-side five modes (Part 08) = the full production RAG closure.

Read more EN
AI Jun 10, 2026

From RAG to Enterprise-Grade RAG Part 08 | One /ask endpoint, how do you know whether this query deserves a faithfulness check

Part 07 wired up four production-grade capabilities — faithfulness, citation, Langfuse tracing, cost tracking — but running them on every query blew latency and cost. This post bundles the four into five query modes (fast / safe / deep_eval / creative / agentic) and routes automatically using six query features (cheap / medium / LLM-fallback). One /ask endpoint, dynamic per query. Concrete workloads: Q3 contract goes to safe, FAQ goes to fast, debug goes to deep_eval, summarisation goes to creative, multi-step tasks go to agentic. Five failure scenarios (router sends contract to fast / Step 6 default flipped to fast / fast gets rerank bolted on / deep_eval opened on user path / traffic spike triggers mode consolidation) and the guardrails for each. Routing concept aligns with LangChain RouterChain's official pattern and LlamaIndex's RouterQueryEngine workflow, not a new invention.

Read more EN
AI Jun 10, 2026

From RAG to Enterprise-Grade RAG Part 07 | Where Production RAG Dies: Measurability, Debuggability, Controllability

Demo-stage RAG looks reasonable, but production-grade RAG dies from answers that look correct but aren't. This article unpacks 3 production-grade capabilities — Measurable (faithfulness / citation check + RAGAS), Debuggable (Langfuse query-level + stage-level tracing), Controllable (deterministic-first + async evaluator + cost tracking) — with a Step 1-7 build order, 6-item production launch checklist, and 5 query modes bridging to Part 08. Faithfulness / Citation check + RAGAS offline eval + Langfuse tracing align with industry RAG evaluation standards (RAGAS, Langfuse, DeepEval official docs).

Read more EN
AI Jun 10, 2026

From RAG to Enterprise-Grade RAG Part 06 | Why Retrieval Decides RAG Quality More Than the Model Does

Vector search finds semantic neighbours, not answers. This piece opens with a concrete failure case, then breaks down the four retrieval-layer upgrades (hybrid retrieval, reranking, parent expansion, citation assembly) and the specific trade-offs they carry across quality, cost and explainability. Each retrieval layer still has clear marginal upside; upgrading the LLM does not — and the budget line that follows from that.

Read more EN
See all posts

Explore by topic

Browse the archive by category

A structured path into essays, notes, and observations across the ideas that shape Daniel’s work.

AI OpenClaw PM Startup Business Web3 Travel Murmurs