simplefloRScientia
This application has been tested to work on Chrome and Firefox browsers only and may not work on other browsers as of now.

Scientia Docs

Scientia is designed for fast iteration. The goal is to get you to a grounded, explainable state without requiring a giant framework.

Open LaboratoryRead Quickstart

⚡ Quickstart

5 min read

Start with sources you trust, then change one thing at a time.

  1. Load documents
  2. Ask a baseline question
  3. Inspect context
  4. Change one variable
  5. Re-run and compare
  1. Load documents you trust. Start with sources where you recognize good retrieval.
  2. Ask a question with a known "good answer." You are establishing a baseline.
  3. Inspect the retrieved context. Look for wrong sections or shallow snippets.
  4. Change one variable. Chunk size, overlap, top-k, or rerank strategy.
  5. Re-run and compare. A/B turns "I think" into "here is the evidence."

🧭 Modes, when to use them

Each mode serves a different part of the RAG loop.

Baseline mode

Quick feedback with minimal setup.

Good for sanity checks, demos, and early exploration.

A/B testing

Isolate one change and compare outputs.

If you are making a decision, use A/B.

Graph mode

Multi-part questions that need a chain of evidence.

Useful for policy or "why/how/compare" prompts.

✅ Evals: how to treat RAG like engineering

Evals turn "it seems better" into "we can trust this." Start small and stay consistent.

  • Source alignment: Did the answer cite the right part of the material?
  • Claim support: Are the key claims supported by retrieved text?
  • Hallucination check: Did the answer introduce details not in sources?
Small evals compound into reliable systems over time.

🧩 Common failure modes

  • Semantic drift: Retrieval returns vector-adjacent but irrelevant text.
  • Bad chunking: Key concepts split across boundaries and lose meaning.
  • The confidence trap: Strong models mask weak retrieval.
  • Overfitting: A tweak helps one query and quietly breaks others.
  • Mode mismatch: Multi-hop questions forced into a single-hop setup.

🔎 Reading traces without getting lost

A good trace answers five questions:

  1. What was the question?
  2. What did retrieval fetch?
  3. What was promoted (and why)?
  4. What was ignored?
  5. What evidence supported the final answer?
Trace summary...
Query: "What policy governs vacation carry-over?"
Top retrieval: Handbook section 3.2
Answer cites section 3.2 and 3.3

If your trace does not answer those questions, it is telemetry, not observability.

❓ FAQ

Does Scientia guarantee correctness?

No. It helps you verify and improve. The point is to make correctness auditable, not automatic.

Do I need a vector database to use this?

Not to start. You can experiment quickly, then choose storage and indexing later.

What should I do if results feel unstable?

Start smaller: fewer documents, clearer questions, and one variable change per iteration.