How ghoumbadji/jobmatcher-ai works
This is a single-repo, single-process Streamlit application that runs locally (or on Streamlit Cloud) and provides career/job-fit recommendations based on:
• A questionnaire UI (Streamlit)
• A semantic scoring engine (SentenceTransformers embeddings + weighted scoring)
• An LLM narrative layer (Google Gemini via API) for personalized written guidance
• Supporting NLP preprocessing (NLTK)
So the “system” is not microservices; it’s a modular Python app with a clear runtime workflow and one notable external dependency: Gemini API.
• Runtime flow from UI → preprocessing → embedding/scoring → visualization → LLM recommendation
• Module boundaries aligned to files (main.py, job_matcher.py, preprocessing.py, genai_recommender.py)
• External systems (Gemini API, optionally model download/huggingface cache, NLTK data download)
• Secrets/config (.streamlit/secrets.toml)
Domain / Decision Engine
job_matcher.py (Core scoring engine)
• Owns: job profiles (Data Analyst/Scientist/Engineer/ML Eng), skill categories, weights, similarity calculations, aggregation into final scores.
NLP Utilities
preprocessing.py
• Owns: normalization pipeline (lowercasing, tokenization, lemmatization, stopword handling).
Generative AI Layer
genai_recommender.py
• Owns: constructing prompt, selecting Gemini model, calling Google Generative AI client, parsing/formatting response.
• requirements.txt → technologies used.
• .streamlit/secrets.toml (mentioned in README, not shown in file tree) → config store for GEMINI_API_KEY.
• SentenceTransformer model all-mpnet-base-v2 → may be downloaded/cached at runtime.
• NLTK corpora/resources (stopwords, punkt, wordnet) → may be downloaded/cached.
These are worth showing as “runtime resources” because they affect deployment and first-run behavior.
Scores → main.py → Plotly charts + tables
5. Scores + best-match + rationale inputs → main.py → Gemini recommender (genai_recommender.py)
How to show boundaries
• Container: “JobMatcher AI (Python/Streamlit app)”
• Contains: main.py, job_matcher.py, preprocessing.py, genai_recommender.py
• External container: “Google Cloud / Gemini API”
• Local runtime resources: model cache + NLTK datasets
Extra details worth annotating (because they affect architecture)
• Model selection: SentenceTransformer all-mpnet-base-v2 (embedding generation)
• Gemini model variants: “2.5 Flash or 2.0 Flash” (dynamic loading)
• Privacy: responses not stored; only sent to Gemini for generation
• Deployment: local/Streamlit Cloud; secrets via .streamlit/secrets.toml
That combination captures both *what it is* (modules + boundaries) and *how it works* (the user-to-output pipeline).