feat: score threshold filtering, .txt support, API key trimming

[biswajeetdev]

What

Three quality-of-life fixes found while doing a thorough test pass:

Fix 1 — Low-relevance chunks passed to LLM (rag.py)

retrieve() returned all TOP_K=4 chunks regardless of their cosine similarity score. When a user asks something not covered by the document, FAISS still returns the closest matches even if their scores are 0.05–0.12 — far below useful similarity. The LLM then generates a hallucinated answer from those irrelevant chunks instead of saying "not in the document."

Added MIN_SCORE = 0.20 floor. If all retrieved chunks are below threshold, retrieve() returns [], and generate_answer already handles empty retrieval gracefully with:

"No relevant content found in the document for that question."

Fix 2 — .txt files not supported (app.py)

The uploader accepted only type="pdf". Plain text files (notes, logs, transcripts, Markdown exports) are a primary use case for a doc QA tool. Added type=["pdf", "txt"] and a branch for UTF-8 decoding via .read().decode() instead of PdfReader.

Fix 3 — Whitespace in pasted API keys causes silent failure (app.py)

Keys copied from browser often include a trailing newline or space. make_chat_client(api_key) creates a client with the whitespace-padded key, which fails on first API call with a cryptic 401. Added .strip() immediately after each st.text_input.

Test

# Install deps
pip install -r requirements.txt

# Run
streamlit run app.py

# Test low-relevance filtering:
# Upload any PDF, ask a completely unrelated question
# → should get "No relevant content found..." not a hallucinated answer

# Test .txt upload:
# Upload any .txt file, ask questions about its content

# Test key trimming:
# Paste a Groq key with leading/trailing spaces — should still authenticate