ais coding ais

my first project(if you could call it that) that i've truly vibecoded initially meant to just be better than my ap csp practice create task turned into a monster

local/offline-ish chatbot with fake confidence but real python files (does more than expected and less than promised)

capabilities

• math parsing
• sympy solving (single var, multi var, systems)
• intent/emotion detection
• context-ref matching
• does the cool ai math thing with words -> vectorization
• topic extraction
• follow-ups
• looks stuff up on wikipedia when local brain is cooked

recent additions:

• correction handling (no, i meant x)
• pronoun/reference repair (can you explain that?)
• light knowledge graph expansion
• adaptive depth (simple vs deep)
• strategy-based replies (direct, socratic, analogy, hints)
• nonchalant style pass (less support bot, more chill)
• no-key wikipedia fallback when local brain fails
• solve-for-x -> solve-for-anything (x, y, a, systems, roots)

limitations

• just about everything

more specifically:

• not a real llm
• not trained on your private brain notes
• regex can and will occasionally do regex things
• memory is short and selective, like me on precalc tests
• confidence score != actual correctness
• symbolic solving can get cursed-looking exact forms for some equations

project structure

• chatbot.py: main loop + glue
• intent.py: intent + entities + math parser
• responses.py: reply gen + tone + strategy + graph use
• memory.py: chat memory + profile-ish state + refs
• knowledge.json: tiny local facts + concept graph
• utils.py: typing effect + text/vector helpers

how it works (high level)

1. user says something
2. intent detector scores it (semantic + pattern + structure + state)
3. entities are extracted (topic, expression, correction target, pronoun ref)
4. memory stores both user and bot messages w/ vectors
5. response generator picks a template, depth, tone, and strategy
6. you get an answer that is sometimes surprisingly solid

quick start

cd chatbot-ollama
.venv\Scripts\python.exe chatbot.py

if you dont have the venv yet:

python -m venv .venv
.venv\Scripts\python.exe -m pip install sympy
.venv\Scripts\python.exe chatbot.py

to stop the venv:

deactivate

for mac/linux equivalent:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install sympy
python chatbot.py

dependencies

• python 3.10+
• sympy (symbolic solving)
• standard library only for everything else (including wikipedia lookup)

example prompts

• explain machine learning in simple terms
• no, i meant artificial intelligence
• can you explain that?
• what is 12 mod 5
• i'm stuck, give me a hint
• explain python with an analogy
• solve for a: a^2 - 5a + 6 = 0
• solve for x and y: 2x + y = 5, x - y = 1
• solve x^3 - 6x^2 + 11x - 6 = 0
• who is alan turing (forces lookup path if local logic taps out)

known weird behavior

• if input is vague enough, it asks for clarification (better than hallucinating ig)
• sometimes it gets the right intent for the wrong reason
• math handling is safe-ish and p strict
• internet lookup depends on internet existing and wikipedia not being mad at your request
• some symbolic answers are exact but unreadable unless you're into math masochism

future roadmap (if i keep maintaining)

• better multi-step tracking
• less brittle parsing
• stronger topic graph + retrieval
• cleaner tests so future me doesnt suffer
• inequality solving and domain constraints (real-only, integer-only)
• optional step-by-step algebra traces

why this exists

because building things is fun and overengineering random stuff is a hobby

contribution status

my bad decisions and several ai models tag-teaming the codebase

used:

• gpt-5.4 for planning
• deepseek-v3.1:671b for initial setup
• combination of opus-4.6, gpt-5.3-codex, and gpt 5.4 for development