Sabitlenmiş Tweet
Arish Ai
1.8K posts

Arish Ai
@Minarul223093
AI & Tech Creator | Tools, News & Reviews Helping creators grow with AI DM for Collaborations 📩
Katılım Nisan 2026
1.4K Takip Edilen966 Takipçiler
Arish Ai retweetledi

Most RAG systems fail in production for one reason: people treat "retrieval + LLM" as the whole architecture.
It isn't. It's maybe 20% of it.
A robust RAG system is a pipeline of decisions, and every stage is a place where quality leaks out before the answer ever reaches the user.
Here's the full picture, stage by stage.
𝗤𝘂𝗲𝗿𝘆 𝗖𝗼𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻
Before you retrieve anything, you translate the question into the language of your data store.
→ Relational data needs text-to-SQL
→ Graph data needs text-to-Cypher
→ Vector data needs a clean semantic query
The store decides the translation. Skip this and you retrieve noise.
𝗤𝘂𝗲𝗿𝘆 𝗧𝗿𝗮𝗻𝘀𝗹𝗮𝘁𝗶𝗼𝗻 (𝗥𝗔𝗚 𝗧𝘆𝗽𝗲𝘀)
One user question is rarely the best question to search with.
→ Multi-Query and RAG-Fusion widen the net
→ HyDE generates a hypothetical answer to search against
→ Decomposition breaks complex questions into sub-questions
The goal is the same: give retrieval a better shot.
𝗥𝗼𝘂𝘁𝗶𝗻𝗴
Now decide where the question should go.
→ Logical routing picks the right data source (graph vs relational vs vector)
→ Semantic routing picks the right prompt for the job
A question about relationships goes to the graph. A factual lookup goes elsewhere. Routing is what makes the system feel intelligent instead of brute-force.
𝗜𝗻𝗱𝗲𝘅𝗶𝗻𝗴
This is the quietest stage and the one that decides everything downstream.
→ Semantic splitting chunks by meaning, not character count
→ Multi-representation indexing stores summaries for retrieval, full docs for context
→ Special embeddings like ColBERT match at the token level
→ Hierarchical indexing (RAPTOR) clusters and summarizes recursively
Bad chunks cannot be rescued by a good model.
𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹
Getting documents back is not the same as getting the right ones in the right order.
→ Refinement cleans and compresses what came back
→ Reranking reorders results by true relevance, not just similarity score
Top-k similarity is a starting point, not the answer.
𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻
The model can now decide it needs more.
→ Active retrieval lets the system fetch again mid-generation
→ Self-RAG critiques its own output and re-grounds it
→ Retrieve-Rewrite-Read loops tighten the answer
Generation becomes a feedback loop, not a single pass.
𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻
None of this matters if you can't measure it.
→ Ragas, Grouse, and DeepEval score faithfulness, relevance, and groundedness
If you're shipping RAG without evals, you're shipping vibes.
The pattern across all seven stages is the same: a robust RAG system is mostly the work that happens before and after the model runs. The LLM is the easy part.
If you had to point to the single stage where most teams lose the most quality, where would you put your money: indexing or retrieval?
cc: Brij Kishore pandey

English
Arish Ai retweetledi

13 Powerful AI Tools that can save 30+ hours in every week:
1. ChatGPT.com (solve any problem)
2. PicWish.com (remove backgrounds)
3. Perplexity.ai (research anything)
4. Suno.ai (compose music)
5. Canva.com (design graphics)
6. ElevenLabs.io (clone voices)
7. Grammarly.com (perfect writing)
8. Luma.ai (create 3D models)
9. RecCloud.com (summarize YouTube)
10. Runway.ml (edit videos)
11. Descript.com (edit podcasts)
12. Syllaby.io (create faceless videos)
13. skysnail.io (create viral thumbnails)
Don’t lose this list, it could be incredibly helpful.

English
Arish Ai retweetledi
Arish Ai retweetledi

🚨 Gemini Pro 1 Year Giveaway! 🚨
✨ Gemini 2.5 Pro
💰 1000 AI Credits/Month
🎬 Veo 3 + Flow AI + Whisk
📚 NotebookLM
📩 Gmail & Docs Integration
To Enter:
1️⃣ Follow @HarshBisen143
2️⃣ Like & RT
3️⃣ Comment “ Send ”
⏳ 48 Hours Only.

English
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi

🔥 Best Multivitamins for Men 💪
Boost Energy ⚡ Build Strength 🏋️♂️ Stay Healthy 🛡️
👉 Follow for daily fitness tips!
👉 @Techai55

English
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi

🚨 | AN 18-YEAR-OLD STARTS MAKING MONEY ON TIKTOK WITH CLAUDE IN JUST 20 DAYS.
No equipment, no experience, no money.
Three weeks ago, the account had zero followers.
Today it’s blowing up. The only difference is that he used Claude—but not the way everyone else does.
Here are the 6 prompts below 👇.
English
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi

AI Tool Rankings That Have Lost Their Spotlight
Honestly, tools I don't even bother launching anymore
Ranked without holding back ↓
Tier D (Completely Dormant)
・Notion AI → ChatGPT is sufficient
・Midjourney → Switched to a free alternative
・Perplexity → Gemini is sufficient
・Gamma → Okay just that first time
Tier C (Only cross my mind occasionally) ↓
English
Arish Ai retweetledi
Arish Ai retweetledi
Arish Ai retweetledi























