Ask questions about a PDF and get answers grounded in its content. The pipeline extracts the text, splits it into overlapping chunks, embeds them with a Hugging Face sentence-transformer, stores them in Chroma Cloud, retrieves the most relevant chunks for a question and has Google Gemini answer using only that context.
flowchart LR
PDF["PDF<br/>files/test_pdf.pdf"] -->|PyPDF2| TXT["Raw text"]
TXT -->|"500-char chunks<br/>100-char overlap"| CH["Chunks"]
CH -->|all-MiniLM-L6-v2| EMB["384-dim embeddings"]
EMB --> DB[("Chroma Cloud<br/>collection: pdf_chunks")]
Q["Question"] -->|embed| DB
DB -->|top-k similar chunks| CTX["Context block<br/>with source + chunk index"]
CTX --> LLM["Gemini<br/>(gemini-1.5-flash by default)"]
LLM --> A["Grounded answer"]
| Step | Function in app.py |
What it does |
|---|---|---|
| 1. Extract | extract_text_from_pdf |
Reads every page with PyPDF2 and joins the text. |
| 2. Chunk | chunk_text |
Fixed-size character windows (default 500) with overlap (default 100), so facts that straddle a boundary aren't lost. Validates the sizes. |
| 3. Embed and store | store_chunks_in_chroma |
Embeds the chunks with all-MiniLM-L6-v2 and writes them to Chroma in batches (default 128). Each chunk gets a stable ID and metadata (source, chunk index, size, overlap). The collection is rebuilt on every run, so results never mix with stale data. |
| 4. Retrieve | query_chroma |
Similarity search for the top-k chunks, printed with metadata and a preview. |
| 5. Answer | answer_with_gemini |
Builds a context block of the retrieved chunks (with a soft character budget) and prompts Gemini to answer only from that context. If the answer isn't there, Gemini says so instead of guessing. |
Grounding rules in the prompt
- Use only the provided context.
- Prefer exact numbers and names.
- Point out disagreements between chunks.
- Keep answers concise.
- Reply "I couldn't find that in the provided documents" when the context has no answer.
Python · LangChain (langchain-chroma, langchain-huggingface) · ChromaDB (Chroma Cloud) · Hugging Face
sentence-transformers (all-MiniLM-L6-v2) · Google Gemini (google-generativeai) · PyPDF2 · python-dotenv
git clone https://github.com/striderzz/RAG_Project1.git
cd RAG_Project1
python -m venv .venv
.venv\Scripts\activate # macOS / Linux: source .venv/bin/activate
pip install PyPDF2 chromadb langchain-chroma langchain-huggingface sentence-transformers google-generativeai python-dotenvCopy .env.example to .env and fill in your own values. .env is git-ignored, so never commit it:
| Variable | Required | Purpose |
|---|---|---|
CHROMA_API_KEY |
yes | Chroma Cloud API key |
CHROMA_TENANT |
yes | Chroma Cloud tenant ID |
CHROMA_DATABASE |
no (default RAG) |
Chroma Cloud database name |
GOOGLE_API_KEY |
for answers | Gemini API key from Google AI Studio |
GEMINI_MODEL |
no (default gemini-1.5-flash) |
Gemini model to use |
RAG_QUERY |
no | Question for the retrieval step |
GEMINI_QUERY |
no | Question for Gemini (falls back to RAG_QUERY) |
Index the sample PDF and ask a question:
set RAG_QUERY=What is Yara Lopez's job title and salary?
python app.pyThe run prints the following:
- the number of chunks created;
- the Chroma batches written;
- a sample stored chunk with its embedding size (384);
- the top-3 retrieved chunks with their metadata;
- Gemini's grounded answer.
To use your own document, put the PDF in files/ and change pdf_path at the bottom of app.py. Change chunk_size
and overlap there to tune retrieval.
files/test_pdf.pdf is a synthetic 100-page document, one fictional employee profile per page (name, age, title,
salary, hobbies, height, weight). It was made for testing retrieval. Every name and number in it is dummy data.
- Page-aware or sentence-aware chunking, with page numbers in the metadata for citations.
- A small web UI (FastAPI or Streamlit) for uploading PDFs and chatting.
- Retrieval evaluation (hit rate / MRR) on a labelled question set.
- A local Chroma option (
chromadb.PersistentClient) for offline use.