Gaurav Upreti

Most devs jump into RAG without a clear structure - then wonder why retrieval is inconsistent and outputs are unreliable Here's the structure I follow for every RAG pipeline: 1/ Document Ingestion: load, clean, pre-process, chunk. Get this wrong and everything downstream breaks. 2/ Embedding: convert chunks to vectors. Model choice matters - match it to your domain. 3/ Vector Storage: ChromaDB for dev, Qdrant or pg-vector for production. 4/ Retrieval: embed query, find top-k chunks. Test this independently before touching the LLM. 5/ Re-ranking: raw retrieval isn't always the best retrieval. Re-rank before passing to LLM. 6/ Generation: retrieved context + user query + structured prompt = reliable output. 7/ Evaluation: no eval means no idea if your pipeline actually works. Each step is its own failure point. Fix the step, not the system. Which part of your RAG pipeline gives you the most trouble? #rag #llm

Most "JWT authentication" tutorials get one critical thing wrong: they never expire the access token properly, or they store it somewhere XSS can reach it. Here's how JWT auth actually looks in a production FastAPI app.

@SahilExec Global regexes are designed for iteration, not repeated boolean checks. For isAdmin(), /admin/ is the safer choice than /admin/g.

This regex is reused across multiple requests: isAdmin("admin") returns true, then false, then true again. Why does the same input give different results here?

@system_monarch Choosing a database is really choosing which trade-offs become cheap and which become expensive. Everything else follows from that decision.

Your database picked a side before you wrote a single query. Every database on earth is either read-optimized or write-optimized. There is no third option. The one you chose is taxing every operation you run, and most engineers never learn which tax they're paying.

@SumitM_X Unless you’re on a symmetric fiber connection, download and upload speeds are usually designed to be very different. It’s an infrastructure decision, not a protocol limitation.

As a developer, Have you ever wondered : WHY downloading 1GB is fast but uploading 1GB feels slower?

@haider1 Different bets optimize for different assumptions. It’s still too early to conclude whether scaling language models alone is sufficient or whether richer architectures are necessary.

anthropic and openai may discover a route to AGI through superintelligent LLMs, but if it fails, they'll have to pivot google and demis are pursuing a broader approach that combines LLMs with planning, world models, agents, and robotics atm, it's too early to know which path will succeed

@askalphaxiv One-shot code generation is useful. Turning a paper into something you can modify, test, and break is where the real learning starts.

GPT 5.6 Sol can one-shot convert an arXiv paper into an interactive Marimo notebook! Great for papers best understood hands on (lots of fun examples in interpretability, inference engineering, agent harnesses, benchmarking, and more) Play around with the notebook, inspect the code, or try the same workflow with your own agents below

@paulg We’ve always built on code we didn’t write. AI just moves that boundary one layer higher.

Before vibe coding became a thing, programming was already evolving in that direction. It already increasingly consisted of installing and configuring stuff other people wrote, without reading the source.

@_jaydeepkarale I’d delete it without hesitation. The only cost is that Python recompiles the bytecode on the next run.

nterviewer: Can I delete the __pycache__ folder? Candidate: "No, Python needs it." Interviewer: "Are you sure?"

The engineering mindset applies here too. Measure risks, validate assumptions, and improve continuously instead of optimizing for either blind optimism or blind pessimism.

@Akintola_steve Unless it creates competitive advantage, I’d rather own the business logic than the security surface.

As a backend developer, which would you prefer? Building your own authentication system from scratch, or integrating a third-party authentication provider into your application? I’d love to know your choice and, more importantly, why.

@goyalshaliniuk Prompt engineering optimizes a single interaction. Loop engineering optimizes the entire feedback cycle. That’s the difference between getting an answer and building a reliable system.

Loop Engineering vs Prompt Engineering What's the Difference? Many people think they're the same. They're not. Prompt Engineering gets AI to generate an answer. Loop Engineering gets AI to improve the answer. Here's the difference 👇

Routing only creates value when the available models are meaningfully different. Otherwise you’re optimizing dispatch, not outcomes.

@DanielGlejzner Hiring and day-to-day engineering are drifting apart. The wider that gap becomes, the less predictive interviews become.

Software hiring has become absurd. At work, you’re expected to use AI to offload manual coding and move faster. Then, to get your next contract, you’re asked to code from memory with no assistance. Pass the interview - and you’re expected to use AI again. It has never been this broken.

@akshdeeps_001 Embeddings capture similarity, not truth. Two chunks can be close in vector space yet both be factually wrong. That’s why retrieval quality depends on both good embeddings and good data.

Let's discuss what is Embedding An embedding converts text into a list of numbers — a vector "king" becomes [0.2, 0.8, 0.1, ...] "queen" becomes [0.2, 0.7, 0.1, ...] notice something? similar meaning = similar numbers = close together in vector space this is why: → semantic search finds meaning, not just keywords → recommendation systems find similar content → RAG retrieves relevant documents "king" - "man" + "woman" = "queen" in vector space language literally has geometry meaning has coordinates and once you understand this — how LLMs, RAG pipelines, and search actually work suddenly makes complete sense

For production agents, reliability isn’t just generating answers. It’s recognizing uncertainty, verifying intermediate results, and knowing when to ask for help.

@devops_prashant A 12ms ping only proves the network is fast. It says nothing about server processing, frontend rendering, or browser work.

Interviewer: You run ping google.com. Response time: 12ms. You open Google in Chrome. Page loads in 800ms. If ping is 12ms, why does the page take 800ms?

@IamAroke LLMs are strongest when the decision has already been made. System design is about deciding what should be built before generating how to build it.

LLMs are great at generating boilerplate but terrible at system design decisions.

@asmah2107 The model isn’t usually the bottleneck. Moving the right information to the right place at the right time is.

Context management is actually a complex distributed systems problem.

@arpit_bhayani Evaluation is becoming just as important as generation. As LLMs improve, the competitive advantage shifts toward measuring quality consistently rather than guessing it.

New write-up is live, and this time I covered G-Eval. It helps answer one important question: how do you know whether what an LLM generated is apt, correct, and aligned with your requirements? G-Eval is a pretty simple framework that leverages Chain-of-Thought prompting over clearly defined rubrics. Instead of simply asking an LLM to "rate this post" or "is this correct?", it makes the evaluation process more procedural and reproducible. If you bluntly ask an LLM to rate something on a scale of 1-5, it tends to exhibit a bias toward picking certain values, lacks a clear audit trail, and often suffers from a bias toward its own model family. G-Eval solves this neatly. In this blog, I covered G-Eval in detail and showed how to adopt it in your AI workloads to actually measure whether your LLM outputs are getting better. If you are looking for a reliable blueprint for LLM evaluation, this guide should help. Give it a read, and like always, I hope this helps.

@ashoKumar89 A Dockerfile is part of your production architecture. Build caching, image size, security, and reproducibility matter just as much as starting the app.

Would you approve this Dockerfile? It works. What production issues do you see?