← All work

04 / NLP / Embeddings

Public repository

Semantic Banking Assistant

A semantic assistant that uses sentence-transformer embeddings and similarity search to understand Banking77 intents and return conversational responses.

Connected an embedding-based intent workflow to a focused Streamlit interface, turning an NLP experiment into an end-to-end demonstration with thresholded semantic retrieval.

INTENT SPACESIMILARITY

01 / PROBLEM

What the system is trying to solve.

Intent systems based only on exact keywords are brittle. The project explores whether sentence embeddings can provide a compact semantic layer for routing natural banking questions into a fixed intent taxonomy.

02 / SYSTEM

How it works.

  1. 01

    Embed Banking77 examples with a sentence-transformer model.

  2. 02

    Compare incoming user queries against intent representations with cosine similarity.

  3. 03

    Apply a confidence threshold and return the strongest semantic matches.

  4. 04

    Cache model resources and keep lightweight session history in Streamlit.

  5. 05

    Expose top-k alternatives so uncertain intent matches are inspectable instead of hidden.

03 / ENGINEERING

Decisions that shaped the project.

01

Banking77 intent taxonomy.

02

Sentence-transformers with all-mpnet-base-v2 embeddings.

03

Cosine-similarity retrieval with thresholding.

04

Top-k intent inspection and conversational session state.

05

Small, deployable Streamlit interface.

04 / EVIDENCE

A few concrete signals.

Banking77

Dataset

all-mpnet-base-v2

Embedding model

Cosine similarity

Retrieval

Streamlit

Interface
  • Banking77
  • Sentence Transformers
  • Semantic Search
  • Streamlit