InferLens

Ask any PDF a question. Get a grounded answer with the source passages cited.

Live demo: inferlens.latentaxis.io

A Retrieval-Augmented Generation (RAG) app that started as an academic capstone and grew into a deployed product. Upload a PDF, ask natural-language questions, and InferLens returns answers backed by semantic search over the document, with the exact retrieved chunks shown beneath each answer.

Try it in 30 seconds

The live demo at inferlens.latentaxis.io ships with three pre-loaded sample documents (a lease, a research paper, a contract) so you can ask questions immediately without uploading anything. Each sample comes with three suggested questions a real evaluator would ask.

Architecture

PDF Upload
   │
   ▼
Text Extraction  (PyMuPDF)
   │
   ▼
Chunking         (~500-word overlapping windows)
   │
   ▼
Embedding        (all-MiniLM-L6-v2 via sentence-transformers)
   │
   ▼
FAISS IndexFlatL2  ◄──── User Query (embedded with same model)
   │                              │
   └──── Top-5 Nearest Chunks ────┘
                │
                ▼
       Anthropic Claude (RAG prompt with citations)
                │
                ▼
   Grounded Answer + Source passages + Share link

Embeddings stay in memory per session. SQLite handles analytics, rate limiting, and 30-day shared-document storage. Deployed on Render with a persistent disk mounted at /var/data for the SQLite database.

What's in here

File	What it does
app.py	Streamlit UI, session state, pipeline orchestration
ingestion.py	PDF text extraction, chunking
embeddings.py	Sentence-BERT model + FAISS index construction
retrieval.py	Query embedding + top-K nearest-chunk search
generation.py	Claude API calls (Q&A + document summarization)
analytics.py	SQLite query log, session table, shared-doc store, monthly rate limit
samples.py	Three pre-loaded demo PDFs generated at runtime via PyMuPDF
sharing.py	Encode/decode shareable answer links (base64-JSON tokens)
Dockerfile / render.yaml	Container build + Render deployment config

Run it locally

git clone https://github.com/koreydillon/inferlens.git
cd inferlens
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # add your ANTHROPIC_API_KEY
streamlit run app.py

Open http://localhost:8501.

Tech stack

Layer	Choice	Why
Frontend	Streamlit	Fastest path from notebook to deployable web app
Embedding model	all-MiniLM-L6-v2 (sentence-transformers)	Tiny (~80MB), fast on CPU, great quality-per-byte for English
Vector index	FAISS IndexFlatL2	Exact search; per-document indexes are small enough to skip approximation
Generator	Claude (claude-sonnet-4-5-20250929)	Strong instruction-following + grounded answer behavior
Storage	SQLite	Single-file, zero-ops, plenty for analytics + share-link TTL
Hosting	Render (Docker)	Persistent disk for SQLite, predictable cold starts

What I learned

• Chunking dominates. 500-word overlapping windows beat both smaller (200) and larger (1000) windows on retrieval quality for legal/research PDFs. Smaller chunks fragmented multi-paragraph reasoning; larger chunks diluted relevance.
• Show the receipts. Surfacing the retrieved chunks under each answer turned the app from "trust me" to "verify it." Completely changed how testers used it.
• Ship the samples. A blank "upload your PDF" screen has bounce. Three pre-loaded documents with suggested questions cut time-to-aha to under a minute.
• Rate limit by email, not IP. Cloud IPs are shared (everyone behind a corporate NAT looks like one user). Email-cookie limits gave a much better signal at minimal friction.

Roadmap

• Hybrid retrieval (BM25 + dense) for queries that lean on rare terms
• Per-document index persistence so re-uploads don't re-embed
• Streaming answers (currently buffered)
• Multi-PDF cross-document Q&A

License

MIT. See LICENSE.

---

Built by Korey Dillon at LatentAxis.