Build a Generative AI Application with Python

Quick Start

A viewer asks, “What can our family watch in Tamil within 90 minutes?” A general LLM can write a convincing recommendation, but it may choose a title that is unavailable, too long or unsuitable for the requested rating.

This Byte separates two jobs that beginners often mix together. The LLM proposes a readable answer; the Python application decides whether that proposal matches approved catalogue facts. You will deliberately test one valid ID and one invented ID to see the boundary clearly.

FROM MODEL CALL TO APPLICATION

Byte project

Build a grounded Movie Discovery Assistant that returns one verified title, a reason, suitability, confidence and any missing information.

PYTHON LEARNING LAB 03

Separate the model proposal from the application decision

The LLM may write a confident answer. Python compares its title ID and constraints with approved catalogue evidence before the user sees it.

CODE → BEHAVIOUR → OUTPUT

Viewer needTamil · U rating · ≤ 90 minApproved cataloguemv-101 · 88 min · Tamil · U

LLM PROPOSAL

{
  "title_id": "mv-101",
  "confidence": 0.94
}

DETERMINISTIC PYTHON CHECK

if title_id not in catalogue:
  reject("Unknown title")
elif runtime > 90:
  reject("Too long")
else:
  approve()

USER-FACING RESULTWaiting for Python

Model confidence is not proof.

Start from Basics

Term	Purpose
Model	Generates or transforms language
System instruction	Stable role, limits and output rules
User instruction	The current request
Context	Verified facts needed for this request
Structured output	Response fields your code can validate
Grounding	Connecting generation to approved evidence
Hallucination	Unsupported or invented output

An LLM call produces text. An application adds validated input, authorised context, output contracts, privacy controls, failure handling and a user experience.

Core Explanation

PYTHON

01from pydantic import BaseModel, Field0203class MovieRecommendation(BaseModel):04    title_id: str05    reason: str06    suitability: str07    confidence: float = Field(ge=0, le=1)08    missing_information: list[str]

CODE RESULTVALIDATED MODEL OBJECT

TRACE THE EXECUTION

EXPECTED OUTPUT

MovieRecommendation(
  title_id="mv-101",
  confidence=0.86,
  missing_information=[]
)

VISUAL EXECUTIONFollow the value through the program

LLM JSONProposed fields

PydanticShape + range checks

Business validationCatalogue truth

The model should return this shape, but Python must still verify that title_id exists in the retrieved catalogue and that rating, language and runtime match the request.

PYTHON

01system_instruction = """02You are a movie discovery assistant.03Recommend only a title present in AVAILABLE_TITLES.04If required viewer information is missing, list it instead of guessing.05Return the requested structured fields.06"""

CODE RESULTMODEL BOUNDARY

TRACE THE EXECUTION

EXPECTED OUTPUT

Allowed: recommend IDs in AVAILABLE_TITLES
Abstain: when required viewer data is missing

VISUAL EXECUTIONFollow the value through the program

System ruleRole + limits

Approved contextCurrent facts

Model proposalNot final truth

Good context is selective. Send the viewer constraints and eligible title facts—not an unlimited chat history or unrelated personal profile.

Validate the proposal against application-owned facts:

PYTHON

01def verify_recommendation(result, catalogue, profile):02    title = catalogue.get(result.title_id)03    if title is None:04        return {"status": "rejected", "reason": "Unknown title"}05    if title["language"] != profile["language"]:06        return {"status": "rejected", "reason": "Language mismatch"}07    if title["runtime_minutes"] > profile["available_minutes"]:08        return {"status": "rejected", "reason": "Too long"}09    return {"status": "accepted", "title": title}

CODE RESULTVERIFIED RECOMMENDATION

TRACE THE EXECUTION

EXPECTED OUTPUT

title_id: mv-101
status: accepted
reason: Tamil · U · 88 minutes
shown_to_user: Yes

VISUAL EXECUTIONFollow the value through the program

Model proposalmv-101

Python checksID + constraints

User resultVerified and shown

Architecture / Flow Diagram

GROUNDED COURSE ADVISORCombine learner context and approved knowledge before generation

FOLLOW THE FLOW

Learner profileSkills + goal

CatalogueVerified courses

LLMStructured proposal

Python validationFinal answer

Remember: The model recommends; Python verifies course identity, prerequisites and required fields.

This is an evidence chain: viewer need → approved facts → model proposal → deterministic Python checks → response. The model never becomes the source of catalogue truth.

Types / Components

Component	Owns	Must not do
Input model	required viewer fields	accept arbitrary unvalidated data
Context builder	selected verified facts	send the entire database
Prompt template	role, task and response rules	mix secrets into instructions
LLM client	provider request	decide business truth
Output model	required response shape	prove factual correctness alone
Business validator	IDs, rating, runtime, availability	trust model confidence
Review boundary	consequential or uncertain decisions	hide missing information

Streaming improves perceived speed for conversational text. For structured decisions, collect and validate the complete object before acting on it.

Tool / Technology Comparison

Pattern	Best use	Main risk
Prompt only	low-risk drafting	no dependable factual boundary
Prompt + structured output	machine-readable results	valid shape may still contain false facts
RAG / retrieval grounding	current approved knowledge	poor retrieval produces weak evidence
Fine-tuning	stable style or repeated behaviour	not a live knowledge database
Deterministic Python rule	exact business constraint	cannot understand every natural-language variation

Strong applications combine model flexibility with deterministic checks rather than forcing one technique to do every job.

Real-World Examples

BEFORE → ENGINEERING ACTION → VISIBLE RESULTFollow the problem until the team can prove the result.

Family recommendation

Before: A model suggests a title unavailable in India.
Engineering action: Ground it with the current regional catalogue.
Evidence after: Python displays only a verified title.

RESULT VERIFIED

Age suitability

Before: A fluent answer ignores the requested rating.
Engineering action: Validate rating and duration deterministically.
Evidence after: Unsafe or unsuitable output is rejected.

RESULT VERIFIED

Missing preference

Before: The viewer gives no language or genre.
Engineering action: Return missing_information instead of guessing.
Evidence after: The interface asks one useful follow-up.

RESULT VERIFIED

These scenarios describe observable application behaviour—not private claims about any company’s internal systems.

The same pattern supports a product assistant grounded in current stock, an HR assistant grounded in approved policy, and a support assistant grounded in published help content. Always keep source identity and access rules visible.

Imagine Pannunga

SEE IT IN PRACTICE

Brief a knowledgeable assistant with the correct file

“Suggest something” gives an assistant too much room to guess. “Choose one Tamil family title under 90 minutes from this approved list” provides a decision boundary and evidence.

Simple Tanglish: “Edhaavadhu suggest pannunga” romba vague. Language, rating, available time, verified catalogue details kuduthaa recommendation useful-ah irukkum; Python final checks pannum.

Use Cases

WHO USES THIS — AND WHAT DO THEY OWN?Connect each concept to an engineering responsibility.

GenAI engineer

Assemble grounded prompts and structured output

Clear ownership

Backend developer

Own validation and business rules

Clear ownership

Responsible AI reviewer

Define abstention and human review

Clear ownership

AI operations engineer

Track quality, latency and token cost

Clear ownership

One Python AI application is a team system: code, quality, security and operations work together.

Practical project

Use a small verified catalogue with title_id, language, rating, genres, runtime and region. Retrieve eligible entries, request a structured recommendation, validate the identifier and constraints, and return missing information instead of guessing.

Track model name, prompt version, input/output tokens, latency and estimated cost. Do not log raw private conversations by default.

Key Takeaways

REMEMBER THIS

Key takeaways

A production GenAI feature is more than one model call.
System instruction, user instruction and verified context have different jobs.
Grounding gives the model relevant approved evidence.
Structured output enables parsing but does not prove factual correctness.
Python must validate identifiers and deterministic business rules.
Missing information should trigger clarification, not confident guessing.
Streaming changes delivery timing, not the trust boundary.
Privacy, cost tracking and human review belong in the initial design.

Final Thought + Next Path

Let the model handle language; let the application own truth, permissions and evidence. You now have a grounded feature that can explain one recommendation without inventing catalogue facts. Next, you will give the application bounded tools and learn how an agent chooses, acts, observes and stops.

FAQ + Knowledge Check

LESSON CHECKPOINTConfirm the concept before moving forward

Choose an answer, inspect the explanation and explain the idea in your own words.

RETENTION

Is structured output automatically correct? No. It can match the schema while containing an unknown ID or unsuitable value.

Is RAG the same as fine-tuning? No. RAG supplies retrieved evidence at request time; fine-tuning changes model behaviour from training examples.

Should the whole conversation always be sent? No. Send only useful, permitted and appropriately retained context.

When is human review required? When uncertainty or consequence is high—for example paid, shared-profile or sensitive decisions.

Who verifies title availability? A) Model confidence B) Python against catalogue C) UI colour
What should happen when language is missing? A) Guess B) Ask or return missing information C) Invent
Does a valid schema prove truth? A) Yes B) No C) Only with JSON
What should grounding contain? A) Relevant approved evidence B) Every record C) Secrets
What should be tracked? A) Only output text B) Model, prompt, tokens, latency and cost C) Nothing

Answers: 1-B, 2-B, 3-B, 4-A, 5-B. Score 4/5 or better before continuing.

Build a Generative AI Application with Python

A movie desk using a real catalogue

Quick Start

Byte project

Separate the model proposal from the application decision

Start from Basics

Core Explanation

Architecture / Flow Diagram

Types / Components

Tool / Technology Comparison

Real-World Examples

Imagine Pannunga

Brief a knowledgeable assistant with the correct file

Use Cases

Practical project

Key Takeaways

Key takeaways

Final Thought + Next Path

FAQ + Knowledge Check

Primary sources

Continue by concept