Ollama Learning Projects

Small projects built while learning to build AI apps on top of a local LLM runtime — Ollama. Every model runs on the machine; there are no cloud API keys anywhere in this repo.

The repo has two tracks, built roughly in this order:

1. Web apps (ai_*/) — FastAPI backends serving a plain HTML/JS frontend
2. Agents (AI Agents/) — LangChain + Streamlit, adding memory, voice, and retrieval

---

Track 1 — FastAPI web apps

Each folder is a self-contained app: app.py (backend) + static/ (frontend).

Project	What it is	Model	What it taught
intro_mistral.py	10-line script hitting the Ollama API	mistral	The raw /api/generate endpoint
ai_chatbot/	Chat page	mistral	Serving a frontend from FastAPI, POST endpoints
ai_text_summarizer/	Paste text → summary	mistral	Form(...) handling, prompt building
ai_code_assistant/	Generate or debug code	codellama	Branching prompts off a mode field
ai_workspace/	Chat + summarize, properly structured	mistral	Routers, Pydantic schemas, async httpx, .env config, SSE streaming
ai_legal_analyzer/	Extract clauses, risks, obligations	phi	Domain-specific prompting
ai_proofreader/	Grammar & spelling fixes	deepseek-r1	Swapping models per task
ai_content_writer/	Article from a topic + style	llama3	Multi-field prompts
customer_support_chatbot/	Support Q&A page	qwq	Persona prompting, surfacing Ollama errors properly
ecommerce_ai_recommender/	Preferences → product picks	granite3.2	Grounding a prompt in a real catalog instead of letting the model invent one
medical_ai_symptom_checker/	Symptoms → general info	medllama2	Where a safety disclaimer belongs: the page, not the prompt
ai_virtual_assistant/	Chat + task scheduling	llama2	Model-driven intent extraction ("format": "json") over keyword matching

ai_workspace/ is the one to read. It is the Milestone 1 rewrite of the earlier apps and fixes what they got wrong — blocking requests calls, copy-pasted Ollama code, hardcoded config, fake 200s on failure, innerHTML XSS. See ai_workspace/README.md for the full before/after.

API surface of the older apps

Project	Endpoint	Body	Returns
ai_chatbot	POST /chat	?prompt= (query param)	{"response"}
ai_text_summarizer	POST /summarize	form: text	{"summary"}
ai_code_assistant	POST /generate_code	form: prompt, mode (generate|debug)	{"code"}
ai_legal_analyzer	POST /analyze_legal_text	form: text	{"insights"}
ai_proofreader	POST /proofread	form: text	{"corrected_text"}
ai_content_writer	POST /generate	form: topic, style	(see Known gaps)
customer_support_chatbot	POST /chat	form: user_query	{"response"}
ecommerce_ai_recommender	POST /recommend	form: preferences	{"recommendations"}
medical_ai_symptom_checker	POST /analyze_symptoms	form: symptoms	{"response"}
ai_virtual_assistant	POST /chat	form: user_query	{"response", "tasks"}

ai_workspace has its own — streaming chat, model listing — documented in its README.

---

Track 2 — LangChain agents (AI Agents/)

Streamlit UIs rather than hand-written HTML, and LangChain instead of raw HTTP.

Day	Script	What it does	Concepts
1	basic_ai_agent.py	Chatbot that remembers the conversation	OllamaLLM, PromptTemplate, ChatMessageHistory, st.session_state
2	ai_voice_assistant.py	Same agent, spoken — CLI loop	speech_recognition (mic in), pyttsx3 (speech out)
2	ai_voice_assistant_ui.py	Voice assistant with a Streamlit UI	Push-to-talk button, persisted history
3	ai_web_scraper.py	URL → scrape <p> tags → summary	requests + BeautifulSoup, truncating context
3	ai_web_scrapper_faiss.py	Scrape a site, then ask questions about it	Chunking, embeddings, FAISS vector search, RAG

The day-1 file keeps its earlier CLI-only versions commented out at the bottom, so the progression (plain LLM → memory → web UI) is visible in one file.

The voice assistant's recognition step calls Google's speech API (recognizer.recognize_google), so that part needs internet. The LLM itself stays local.

---

Setup

1. Ollama and models

Install Ollama, then pull the models the projects use:

ollama pull mistral      # chatbot, summarizer, workspace, agents
ollama pull llama3       # content writer, day-1 agent
ollama pull codellama    # code assistant
ollama pull phi          # legal analyzer
ollama pull deepseek-r1  # proofreader
ollama pull qwq          # customer support chatbot
ollama pull granite3.2   # e-commerce recommender
ollama pull medllama2    # medical symptom checker
ollama pull llama2       # virtual assistant

Any of the four newest apps will also run against a model you already have — set OLLAMA_MODEL instead of pulling another few gigabytes:

OLLAMA_MODEL=mistral uvicorn app:app --reload    # macOS / Linux
$env:OLLAMA_MODEL="mistral"; uvicorn app:app --reload   # PowerShell

Only pull what you need — each is a multi-GB download. Ollama serves on http://localhost:11434; check it with ollama list.

2. Python environment

python -m venv ollama_env
ollama_env\Scripts\activate      # Windows
source ollama_env/bin/activate   # macOS / Linux
pip install -r requirements.txt

requirements.txt covers Track 1 only. The agents need more:

pip install streamlit langchain langchain-community langchain-ollama \
            langchain-huggingface sentence-transformers faiss-cpu \
            beautifulsoup4 numpy SpeechRecognition pyttsx3 pyaudio

pyaudio (microphone access, day 2) needs a system build toolchain and is the usual install failure — skip it unless you're running the voice assistant.

---

Running a project

FastAPI apps — run from inside the project folder; static/ and .env are resolved relative to the working directory:

cd ai_code_assistant
uvicorn app:app --reload

Then open http://127.0.0.1:8000 (API docs at /docs). ai_workspace uses a package layout, so it's uvicorn app.main:app --reload instead.

The older apps all bind port 8000 — run one at a time, or pass --port 8001.

The four newest ones each have their own default port so they can run side by side: customer_support_chatbot 8001, ecommerce_ai_recommender 8002, medical_ai_symptom_checker 8003, ai_virtual_assistant 8004. That default lives in each app's python app.py block, and uvicorn does not read it — so pick one:

cd customer_support_chatbot
python app.py                           # uses the port baked into app.py
uvicorn app:app --reload --port 8001    # uvicorn needs the port spelled out

Streamlit agents — run from the repo root:

streamlit run "AI Agents/day1/basic_ai_agent.py"

Opens on http://localhost:8501.

---

Repo layout

intro_mistral.py             first contact with the Ollama API
ai_chatbot/                  ┐
ai_text_summarizer/          │
ai_code_assistant/           │
ai_legal_analyzer/           ├─ Track 1: FastAPI + static frontend
ai_proofreader/              │  (app.py + static/index.html each)
ai_content_writer/           │
customer_support_chatbot/    │
ecommerce_ai_recommender/    │
medical_ai_symptom_checker/  │
ai_virtual_assistant/        ┘
ai_workspace/                the structured rewrite — app/, routers/, config, .env
AI Agents/day1..day3/        Track 2: LangChain + Streamlit
requirements.txt             Track 1 dependencies
ollama_env/                  virtualenv (gitignored)

---

Known gaps

Kept honest rather than quietly patched — this is a learning repo, and these are the next things to fix.

• ai_content_writer is broken. generate_content() builds the prompt and calls Ollama but never returns the result, so /generate responds null while the page reads data.content. Its static/script.js is empty too — the real logic is inlined in index.html.
• Empty placeholder script.js files in ai_chatbot, ai_code_assistant, and ai_content_writer; those pages use inline <script> blocks. (The four newest apps had the same gap and no longer do.)
• ai_text_summarizer sends "Mistral" (capitalised) as the model name instead of the MODEL_NAME constant right above it.
• Track 1 apps other than ai_workspace share the same weaknesses: blocking requests calls, no timeouts, hardcoded model and URL, duplicated Ollama plumbing. ai_workspace exists because of them. The four newest apps have since had the worst of that fixed — they check Ollama's status code, time out, and read OLLAMA_MODEL — but they are still blocking, and each still carries its own copy of the same Ollama code. Folding them onto the ai_workspace structure is the real fix and hasn't been done.
• ai_virtual_assistant forgets everything on restart. scheduled_tasks is a plain in-process list: it is wiped when the server restarts and is shared by every visitor, since there are no user accounts. It needs a database.
• Scheduled due times are never parsed. The assistant stores the due time as the free text the model extracted ("tomorrow at 5"), not a real datetime, so nothing can sort, remind, or expire.
• Agent dependencies aren't pinned anywhere — Track 2 has no requirements.txt of its own.