BM25 / keyword search
Like a book index: excellent for product codes, names and rare exact terms. It scores documents using term frequency while reducing the influence of very common words.
01Plain words
Every idea starts with the shortest useful explanation.
02Moving pictures
Each diagram shows what the concept actually does.
03Familiar comparisons
Practical examples and familiar analogies make the ideas easier to remember.
Watch & explore
Fresh from my channel.
Videos from rono mondal, right here in Oasis.
Loading the latest videos…
Interesting playlists from my channel.
Pick a playlist and keep exploring on YouTube.
How the ideas fit together
One family. Different roles.
AI is the broad field. It includes learned systems and explicit rules, logic, planning and search. Machine learning learns from data; neural networks are one family of ML models; deep learning uses neural networks with many layers.
Generative AI describes a capability. It creates content and is not another nesting level. Many prominent generators use deep learning, but generative modelling also includes other approaches.
Modern LLMs are deep neural networks. Generative LLMs produce text token by token. An agent combines a model with instructions, tools, state and a control loop; it is not simply a larger LLM.
Choose your depth
One reference.
Three ways in.
Follow the whole path or enter where you are. Every advanced idea links back to the foundations it depends on.
The RAG field guide
From a document
to a grounded answer.
RAG is a relay race. Each stage hands a smaller, better-labelled bundle to the next. A brilliant final runner cannot recover evidence dropped near the start.
→
→
→
→
→
Load: collect authoritative sources, preserve permissions and record where every item came from.
Vector / semantic search
Like asking a librarian who understands paraphrases: it can connect “locked out” with “account recovery,” even when the words differ.
Hybrid search
Like consulting the index and the librarian, then merging their lists. Reranking gives the final candidates a slower, closer read.
The RAG vocabulary map
Every important handoff,
under one umbrella.
Follow left to right, then open any term for its plain-English lesson, working example, limits, analogy and further reading.
- 01 · PREPAREMake sources usableLike cleaning, labelling and cutting ingredients before cooking.
- 02 · REPRESENTBuild searchable signalsLike preparing both a subject map and a precise back-of-book index.
- 03 · ASK & FINDTurn intent into candidatesLike rephrasing a request, then searching only shelves you may access.
- 04 · COMBINEMerge and refine evidenceLike merging judges’ lists before giving finalists a closer review.
- 05 · ANSWER & PROVEFit, cite and testLike writing from highlighted evidence and keeping every footnote.
“Normalization” can mean three different jobs.Text normalization makes equivalent text consistent. Vector normalization gives vectors a standard length for compatible similarity math. Score normalization places unlike retrieval scores on a comparable scale before fusion.
Fast judgment and deliberate reasoning
System 1 decides.
System 2 works it through.
The names began as an analogy for human cognition. In current AI discussions they describe different computational behaviours, not minds inside machines. A system can use either mode or combine both.
Two meanings to keep separate.System 1-like can broadly mean a quick model response. System One Model is also a newer label used by Jev and open projects such as Laya for models built specifically to return bounded, typed decisions. An ordinary fast LLM is not automatically a System One Model.
01
FAST PATHSystem 1
Maps the current state directly to a constrained judgment: choose one option, score candidates, or estimate yes/no probability.
support_ticket→
{ route: "billing", confidence: .94 }02
DELIBERATE PATHSystem 2
Spends extra inference work to decompose, calculate, plan, use tools, inspect evidence, and revise before answering.
complex_request→
plan → search → test → answer| Question | System 1 / typed decision path | System 2 / reasoning path |
|---|---|---|
| Core job | Make a fast, bounded judgment. | Solve an open-ended, multi-step problem. |
| Typical output | A choice, score, probability, route, rank, or accept/reject signal. | An explanation, plan, proof, program, research synthesis, or tool-driven result. |
| Computation | Usually one parallel forward pass with no token-by-token prose generation. | Often autoregressive generation plus reasoning tokens, reflection, search, or tool loops. |
| Latency and cost | Low and predictable enough for routing, filtering, and high-volume decisions. | Higher and more variable because harder tasks receive more computation. |
| Good fit | Intent routing, moderation, triage, lead scoring, policy gates, and candidate ranking. | Mathematics, complex coding, investigation, planning, and decisions that need evidence. |
| Main risk | A crisp probability can hide a bad question, weak calibration, or a domain shift. | A long reasoning trace can still contain false assumptions, tool errors, or unnecessary work. |
| Examples | Jev, Laya, Decision Kai/Lex, Jebadiah, and educational MiniSystemOne. | OpenAI reasoning models, Gemini thinking, Claude extended thinking, and DeepSeek-R1. |
THE PRACTICAL PATTERNFast where confidence is earned. Deliberate where stakes or uncertainty rise.
Request→S1 route→S2 solve→S1 check→Action or human review
Recent System One model ecosystem
Typed answers
instead of prose.
These projects receive context plus a typed question. They return values software can use directly. Capabilities and evaluation methods differ, so benchmark each model on the exact decisions your application makes.
A hosted, closed-weight model from TypeSafe AI and the project that popularised the recent label. Its API exposes choice, score, and noul primitives, returning typed probabilities rather than free-form text. Best suited to software decisions where a fixed answer contract and low latency matter.
Watch for: provider dependence, calibration on your domain, and the fact that a probability is not a guarantee.An Apache-2.0, open-weight family using encoder backbones for single-pass typed decisions. Published checkpoints cover English and multilingual use, with roughly 322M–421M parameters and 512–1,024-token input limits. It offers a Jev-compatible API for local deployment.
Watch for: the project explicitly describes its base checkpoints as foundations to fine-tune; do not assume strong zero-shot performance.Compact open decision-model work from the semantic-router ecosystem. It applies the same typed primitives to local classification and scoring, aiming to replace larger generative calls in narrow decision steps.
Watch for: confirm the task, language, license, and checkpoint-specific evaluation before production use.An open System One-style model based on a larger generative backbone and adapted to emit constrained decision probabilities. It explores the same interface at a different size and architecture point.
Watch for: a larger backbone changes serving cost, and benchmark results across different prompts or datasets are not directly comparable.A small educational implementation trained from scratch to demonstrate parallel, non-autoregressive decisions. It is useful for studying the architecture and training recipe.
Watch for: treat it as a learning and experimentation project unless your own evaluation supports deployment.Agent framework atlas
Choose by control flow,
not by logo.
These libraries overlap, but each starts from a different centre of gravity. You can also build directly with a model SDK and ordinary application code.
DATA-CENTRICLlamaIndexingestion · indexes · retrieval · data agents
→COMPOSITIONLangChainmodels · tools · retrievers · agents
→STATEFUL CONTROLLangGraphnodes · shared state · checkpoints · review
→ROLE-BASED TEAMSCrewAIagents · tasks · crews · structured flows
Start simple. If one model call and two functions solve the task, a large orchestration layer may add more debugging than value.
Prefer workflows for known processes. Use an agent loop when the next step truly depends on evidence discovered at runtime.
Frameworks do not supply safety. Permissions, idempotency, timeouts, evaluation, tracing and human approval remain application responsibilities.
The one idea to keep
AI is powerful.
It is not magic.
It learns patterns, works within limits and needs clear goals, good information and thoughtful checks. Understanding those pieces is how you use it well.
Review the path ↑