1) DAG workflow
a) Pre designed prompts + code paths

b) Modular components [Yes or No prompts]

2) Agents

LLM setups to make their own decision

3) Agentic workflows

Hybrid architecture

LangChain:
It is designed for LLMs. It can be useful for custom workflows, tool integration and agent collaboration. 

LangGraph:
LangGraph library used for build a stateful, multi actors with LLMs

State: a TypeDict that flow through graph and updated by each nodes.

Nodes : Python function used to receive update and return updates

Edges : Connection between Nodes

Reducter : control how state updates combine with over write or accumalation

A RAG is an interactive system (chatbot) that combined a retrieval (static content) with a dynamic conversation generator.

An index : A mechanism to convert a raw data into Vector database

A retriever: Closed tied to the index and retrieve a data from database based on query

A generator : an LLM to reason a through user query and the retrieval knowledge to provide an inline conversational response.

A system has understand of context and meaning of user query and matches against the available document for retrieval. It can find relevant document without having to rely on exact words or n-gram matching. It often uses a pre-trained large language model to understand the nuance of the query and the documents.

We often use of cosine similarity to define a raw data that produce in vector database.

Cosine is bounded between 1 and -1.

smaller angle == Cosine [up]

larger angle == Cosine [down]

Perpendicular == Cosine 0

A dense model uses all the parameter for every tokenized. It would be better for semantic similarity, paraphrases and conceptual queries. [Example GPT3 & GPT4 are using Dense model]

Pros:

* Easier for training the model

* Simple architecture

Cons:

* high computation cost

A sparse model is used for small set of parameter for each token. The most common Sparse architecture is Mixure of Experts.

Pros:

* Less computation cost

* Better scaling efficiency

Cons:

* Potential training

* More complex architecture

Cross Encoders:
A cross Encoder is a non-generative LLM specifically designed to take in two inputs separated by a special token and return a single output.[Ex. BERT]

Reasoning models like Deepseeks R1, OpenAI are orginal "o" series and Anthropic's Claude/Opus 4 models are autoregressive LLMs that have been trained to give a discursive chain-of-thought reasoning step before giving a user response.

For example, the Claude 4 series of LLMs provide separate reasoning tokens alongside the messages to the user, as is common with most frontier reasoning LLMs.

1) Solo Agents

2) Supervisor + specialist agent

1. Encode text into semantic vectors with little/no file tuning

2. Fine tune a pre-trained LLM to perform a very specific task using by Transfer Learning

3. Query an LLM to solve a task which was pre-trained or could intuit.

Two types of LLMs now.

1) Auto encoding LLMs - Learn a entire sequence by predicting tokens (words) given past and future context.   It is best for classification and embedding + retrieval tasks. [Example BERT]

2) Auto regressive LLMs : It will predict a future token 

LLMs excel at task that require reasoning using context and input information in the conjunction to produce a nuanced answer.

AI agents are semi autonomous systems that interact with environment, make decisions and perform tasks on behalf of users.

Autonomy - They can perform tasks without continuous human intervention.

Decision Making - Use data to analyze and choose actions

Adaptability - Learn and improve over time with feedback.

Optimizing Models:

Speculative Decode : Using an assistant model to guide next token perdition

Caching OS models : Implementing prompt caching with open Source models

Quantization : Reducing computation requirement of neural network.

Distillation : Transfer knowledge from large model into small through targeted fine tuning.

Speculative Decoding:

Assistant agent calls for forward method of calling [calling parameter over and over again].  The main model simply verifies which token is agreed with request.

1) Describe - Describe your problem or expectation

2) Define - Stack, database, auth & deployment

3) Generate - Cursor will take care of codes

 

A Kubernetes operator is a specialized controller designed to extend Kubernetes API, enabling the management of complex application through declarative configurations.

Kubernetes Operator operate within continuous reconciliation loop. This cycle begins when user create a resource, It prompting controller to monitoring a change and take a necessary action to ensure a desire state. Operator allow user to defined a desired state of application in custom resources, while operator

controller continue reconcile the actual state with this desire state, embodying the operational expertise of human site reliability engineer.

KubeFlow : It is a primary of orchestration tool for MI workflow. It is focus on training aspect of models.

AI application such as Claude or chatGPT can connect to data sources, tools [search engine] and workflow [prompts] through MCP and perform a tasks.

MCP like an interface which communicated to MCP client and discover their requirement and offer available services for their requirement. 

Embedding represent text as set of numerical data along with tensors (different dimensions)

Each dimension will store some information about text meaning or syntactical meaning.

Each words or sentence with similar meaning are stored near by vector space.

Models will learn to place a similar words or sentences close  together in the embedded space.

Common pre-trained models such as BERT and RoBERTs are  used for generating an embedding inside of vector space.

We can able to use an embedded for NLP tasks like semantic search, text classification and sentimental analysis.

It is integrate an AI agents to enhance the RAG approach. It will breakdown from complex queries into manageable parts and using API tools where need to augment processing and better result.

Traditional LLM model is refine a dataset into pretraining workflow. The pretraining send a data into fine tuning model and give a precise collected output data. It will send it to human feed back and correct incase of any mismatch with fine tuning model.

    DeepSeek applied supervised fine-tuning to refine the models' capabilities. This involved training on datasets containing reasoning and non-reasoning tasks. Notably, reasoning data was generated by specialized "expert models" trained for specific domains such as mathematics, programming, and logic. These expert models were developed through supervised fine-tuning on both original responses and synthetic data generated by internal models like DeepSeek-R1-Lite. The use of expert models allowed DeepSeek to generate high-quality synthetic reasoning data to enhance the primary model's performance.