RAG experiment
A better answer starts with the right knowledge base
AskYou is an educational experiment that makes the path between a question, retrieved context, and a generated answer with demonstrative sources visible.
Context
A convincing answer may still be outside the knowledge base.
The experiment starts from a simple limitation: a model can only use contextual knowledge when that knowledge reaches the prompt. Instead of hiding this step, AskYou makes it the main story of the interface.
- Work with demonstrative data
- Separate retrieval from generation
- Observe the selected sources
Product question
What needs to be retrieved before answering?
The experience organizes a question to pass through vector search before reaching Gemini. The value is giving the model a sufficiently relevant slice of context without pretending the whole base should be sent on every round.
- Represent the question in a vector space
- Apply similarity and a threshold
- Compose the prompt with context and sources
Architecture decision
Retrieval is an observable product step.
The pipeline uses Sentence Transformers for embeddings, FAISS for search, and LangChain with Gemini for synthesis. This separation makes it possible to adjust retrieval without mixing the diagnosis of the base with the quality of the final language.
- Versioned synthetic dataset
- FAISS as a similarity index
- Gemini receives retrieved context
Limits and next steps
Retrieval is visible; production still requires measuring the effect.
In the synthetic sample, 88% of questions retrieved at least one relevant source and grounded answers improved by 24 percentage points over the no-context prompt. AskYou still lacks authentication, observability, and cost control.
- 88% relevant retrieval in the sample
- 24 percentage point gain with explicit context
- Add telemetry and cost control
Conceptual architecture
From question to a grounded answer built on the right context.
The architecture makes embeddings, vector search, retrieved context, and generation visible before the answer.
Question
Person's intent converted into a vector representation.
Embeddings
A model that brings similar questions and documents together.
FAISS
Fast search for the closest sources.
Context
Retrieved excerpts organized for the prompt.
Gemini
A contextualized answer from the demonstrative base.
Featured learning