CV RAG Assistant

Ask questions about a CV and get answers grounded in its contents. The application uses section-tagged chunks, OpenAI embeddings, local persistent Qdrant storage, metadata-filtered retrieval, and a LangChain chat model.

Architecture

1. CV extraction: rag.py reads text from PDF and TXT files using pypdf or UTF-8. It also extracts PDF hyperlink annotations and visible URLs, then adds formatted GitHub, LinkedIn, and project links to the matching section chunks.
2. Section metadata: headings are extracted from the CV before chunking. Each chunk includes its normalized section, original section_header, an expanded section_header_text for matching, source filename, chunk index, CV ID, and stable chunk ID. If headings are missing, lightweight text cues assign best-effort section labels.
3. Chunking: each section is split into overlapping 180-word passages with 35-word overlap.
4. Embeddings and storage: LangChain's OpenAIEmbeddings uses text-embedding-3-small. Vectors and metadata are stored in a persistent local Qdrant database at qdrant_storage/.
5. Pre-retrieval — section header gating: all chunk metadata is loaded into memory. The question and each unique section header are embedded with the same model, cosine similarity scores each header, and only headers above HEADER_SIMILARITY_THRESHOLD (default 0.18) are passed to Qdrant as a metadata filter before vector search. The sidebar lets the user adjust the threshold or further restrict eligible section labels. This was chosen because CVs are organized into meaningful sections, so gating by relevant headings narrows the search without changing the user's question.
6. Post-retrieval — MMR reranking: Qdrant considers up to three times top_k (up to 12 by default) relevant candidates, then maximal marginal relevance (MMR) selects the final passages by balancing query relevance with diversity. This was chosen because CV chunks overlap and neighboring chunks can repeat the same details; a more varied final context gives the answer model broader evidence while keeping passages relevant. The balance is set to 0.5.
7. Answering: LangChain's ChatOpenAI sends the reranked passages and question to the configured OpenAI model. The prompt requires answers to stay within the passages and report missing information. The UI shows passing header scores and final passages with section and source filename.

The local Qdrant database persists CV text, vectors, and metadata on disk. Remove qdrant_storage/ to delete all locally indexed CVs. API keys are read from .env or entered in the sidebar and are not saved by the app.

Models and configuration

The default chat model is gpt-6-luna, selected for cost-sensitive chat usage. The embedding model is text-embedding-3-small, a compact OpenAI embedding model suitable for semantic search. Model names can be changed in .env.

The project contains an empty .env template. Add your key to OPENAI_API_KEY, then adjust the model settings if needed:

OPENAI_API_KEY=your-key
OPENAI_MODEL=gpt-6-luna
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
QDRANT_PATH=qdrant_storage
HEADER_SIMILARITY_THRESHOLD=0.18

Never commit API keys. .env and the local Qdrant database are excluded by .gitignore.

Start on Windows

Double-click start.bat. It creates .venv if needed, installs dependencies, starts Streamlit, and opens the UI in your browser. The first dependency installation requires an internet connection. You can also run it from PowerShell:

cd "D:\curt\Gen AI\task2\CV-Based-RAG-Application"
.\start.bat

Manual setup

Requires Python 3.10 or newer.

py -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
python -m streamlit run app.py

Upload a text-based PDF or TXT CV. Scanned PDFs need OCR before upload. The first time a CV is indexed, its chunks are embedded using the configured OpenAI embedding model and written to local Qdrant. Later app starts reuse that index when the same CV and embedding model are uploaded.

Naive RAG scope

Retrieval uses a metadata header similarity gate before search and MMR reranking after candidate retrieval. It does not use query rewriting, query expansion, HyDE, or contextual compression. Section labels are inferred from common CV headings and simple cues, so review the displayed source passages when a label seems incorrect.