Callisto — Portfolio RAG Knowledge Platform

A demo-scale retrieval-augmented generation (RAG) system for document indexing, hybrid search, and citation-grounded answer assembly.

Recruiter-facing summary

Callisto is a full-stack portfolio project that shows how I design and implement a practical RAG pipeline, from ingestion to retrieval and answer assembly. I built it to be inspectable and honest: the default stack uses deterministic hash-based embeddings, weighted reranking, and heuristic answer synthesis so the whole workflow can run locally without paid model APIs. I am a University of Maryland student studying Information Science and Electrical Engineering with a Business minor.

What this project demonstrates

• Building a complete document QA workflow: ingest → chunk → index → retrieve → synthesize answers with citations.
• Implementing hybrid retrieval patterns that combine lexical and vector signals.
• Structuring a FastAPI backend with explicit service boundaries for ingestion, retrieval, reranking, and answer assembly.
• Designing demo-safe defaults (local embeddings + heuristic synthesis) that can be swapped for real model providers.

Tech stack

• Backend: FastAPI, SQLAlchemy, Pydantic, Alembic
• Frontend: React, Vite
• Retrieval/Data: PostgreSQL, FAISS, lexical retrieval (BM25-style scoring)
• Dev tooling: Docker Compose, Makefile, pytest

Architecture overview

Callisto follows a straightforward RAG flow: documents are uploaded, chunked, embedded/indexed, retrieved through hybrid search, then reordered with weighted reranking before template-based answer assembly.

• Architecture notes: docs/ARCHITECTURE.md
• API surface: docs/API.md

Implementation honesty notes:

• Answers are assembled via heuristic answer synthesis (template-based), not remote LLM generation by default.
• Candidate ordering uses weighted reranking over retrieval features, not cross-encoder reranking.
• Embeddings are deterministic hash-based by default so the app works offline/local-first; you can swap in a real embedding model.

How to run locally

One-command setup

make bootstrap

One-command app start (Docker)

make dev

Then open:

• Frontend: http://localhost:5173
• API docs: http://localhost:8000/docs

Seeded users:

• admin@calisto.ai / password123
• member@calisto.ai / password123
• viewer@calisto.ai / password123

Demo workflow

1. Run make bootstrap (first time) and make dev.
2. Sign in as admin@calisto.ai.
3. Open Documents and upload one of the files from data/samples/ (or paste text content).
4. Open Chat and ask a question tied to that document.
5. Review citations/snippets returned with the answer.
6. (Optional) Use python scripts/evaluate_retrieval.py to run the sample retrieval evaluation set against the local API.

Screenshots / Portfolio Preview

Repository screenshots are listed in docs/screenshots/README.md.

Current screenshots:

• docs/screenshots/dashboard.png
• docs/screenshots/documents.png
• docs/screenshots/chat.png
• docs/screenshots/admin.png
• docs/screenshots/api-docs.png
• docs/screenshots/audit.png
• docs/screenshots/metrics.png

Design/portfolio page:

• docs/preview/index.html

Architecture Decisions

• Chunking strategy: docs/chunking-strategy.md

Limitations and future work

• Default embeddings are deterministic/hash-based, so semantic quality is limited compared with modern embedding APIs.
• Answer synthesis is template-based; integrating a real LLM provider is planned but not required for local demo use.
• Retrieval is tuned for demo-scale local datasets, not large hosted corpora.
• There is no full CI deployment pipeline in this repository today.

Resume bullets

• docs/resume-bullets.md

License

• MIT