Blog

142 total

142 articles across AI, product, startup, business, Web3, and travel.

AI Jun 12, 2026

From RAG to Enterprise-Grade RAG Part 11 | Series Closure: From 5 Query Modes to the 2026 Production RAG Landscape

Parts 02-10 walked through the complete journey from minimum viable RAG to production. This closer turns the series into a coordinate system: how 11 capability flags collapse into 5 query modes, why the learning path has to be re-sorted after one full round, what I deliberately chose not to build, and where the shared production backbone, Agentic RAG, Structured / SQL RAG, Multimodal / Document RAG, Long-context Hybrid, GraphRAG, and Hierarchical RAG / RAPTOR sit in the 2026 Production RAG landscape.

Read more EN
AI Jun 12, 2026

From RAG to Enterprise-Grade RAG Part 10 | Deploying My RAG: 10 Production Pitfalls

Going from a local prototype to a public API is not the final step; it is another engineering review. This article walks through 10 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. It also adds the production request path for agentic/tool workflows: timeout boundaries, worker queues, retries, and observability.

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 collapsed capabilities into five query modes, but strict query-side routing cannot save a broken document side. This post walks through the seven stations a document travels after upload: parsing, raw storage, Postgres metadata, ingestion queue, auth/permission, document APIs, and citation viewer. It also places Multimodal / Document RAG in the right layer, explains what text / table / image / layout signals must survive ingestion, and shows why the Vector DB cannot be the source of truth and ACL cannot rely on one filter.

Read more EN
AI Jun 10, 2026

From RAG to Enterprise-Grade RAG Part 08 | Query Router and Agentic RAG: how one /ask endpoint decides the next step

Part 07 wired up faithfulness, citation, tracing, and cost tracking, but running every capability on every query made latency and cost unsustainable. Part 08 collapses those capabilities into five query modes (fast / safe / deep_eval / creative / agentic) and makes Agentic RAG explicit as a bounded multi-step runtime loop. This post adds two routing layers: a mode router that chooses the capability bundle, and a pattern router that decides whether the query should go to document RAG, SQL / API tools, long-context packing, or an agentic workflow. The point is not free-form agents; it is deterministic features, escape hatches, per-step traces, budgets, timeouts, and final verification around one production /ask endpoint.

Read more EN
AI Jun 10, 2026

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

Demo-stage RAG may look reasonable, but production-grade RAG fails when answers look correct while remaining unsupported. This article turns the Production RAG quality control stack into six layers: retrieval evidence, synthesis, faithfulness check, citation check, tracing, and offline eval. It explains why RAGAS / LLM judges should stay out of the synchronous user path, and adds Agentic RAG per-step trace / per-step check: every step must be replayable, evaluable, and stoppable.

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 starts from a concrete failure case and turns the retrieval layer into an enhancement pipeline: hybrid retrieval improves recall, reranking improves precision, parent expansion + compression improve answerability, and citation assembly improves explainability. It also adds when to use Long-context Hybrid and how to think about the quality / latency / cost trade-off.

Read more EN
AI Apr 2, 2026

ComfyUI Series 04 | Where do you actually get models from? What Civitai and Hugging Face do differently, and how to read a model page properly

Who is this for? Anyone who has ComfyUI running and is now ready to install models, only to discover that one page is full of checkpoints, another is full of LoRAs, and everything has version numbers, strange file names, and far too many download buttons. If you’ve just finished the install, this is the right place to go next.

Read more EN