QueryLens

An AI-powered natural language interface for querying an Oracle database — ask a question in plain English, get back a validated SQL query, the results, a plain-English explanation, and relevant charts. Built during the 48-hour IADS Hackathon (2026).

What it does

• Natural language → SQL — converts a plain-English question into an Oracle SQL SELECT query using OCI Generative AI (Cohere command-r-plus).
• Retrieval-augmented semantic search — for more open-ended, descriptive questions, generates a query embedding (Cohere embed-english-v3) and matches it against pre-computed product embeddings stored directly in Oracle, using a custom cosine-similarity SQL function. This keeps retrieval inside the database rather than standing up a separate vector store.
• SQL safety validation — every generated query is checked before it runs: only SELECT statements are allowed, and any INSERT/UPDATE/DELETE/DROP/ALTER/CREATE/TRUNCATE/MERGE/GRANT/REVOKE is rejected outright.
• Plain-English SQL explanations — breaks down what the generated SQL is doing, clause by clause, for non-technical users.
• Multi-turn chat — keeps conversation context so follow-up questions can build on previous ones.
• Follow-up question suggestions — after a query runs, suggests related questions (grouping, filtering, comparisons) to encourage further exploration.
• In-result visualizations — dynamic charts built from the current query's results, alongside aggregated charts summarizing the full dataset.

My contribution

This was a team project built in a 48-hour hackathon. I worked on the frontend (the Streamlit chat + tabs interface) and the RAG/semantic search implementation — the embedding pipeline and the cosine-similarity retrieval that sits inside the SQL layer.

Architecture

Layer	Choice
Data	Amazon product sales data, Oracle Autonomous Database
SQL generation & explanations	OCI Generative AI — Cohere command-r-plus
Embeddings / semantic search	OCI Generative AI — Cohere embed-english-v3
Frontend	Streamlit
Safety layer	Keyword-based SQL validator, read-only by design

Repository structure

nl-query-agent/
├── README.md
├── requirements.txt
├── .env.example
├── app.py                       # main entry point
├── db.py                        # Oracle connection pool
├── populate_vectors.py          # one-time script: computes & stores product embeddings
├── ai_sql_generator.py          # NL → SQL generation + RAG semantic search
├── validator.py                 # SQL safety validation (read-only enforcement)
├── sql_explanation.py           # plain-English SQL explanations
├── chat_tab.py                  # multi-turn chat interface
├── insight_tab.py               # AI-generated result insights
├── followup_suggestions_tab.py  # follow-up question suggestions
└── charts_in_results_tab.py     # per-query and aggregated charts

Data

Not included in this repo. The app expects an Oracle Autonomous Database with a product/sales table (schema referenced in ai_sql_generator.py) and a PRODUCT_VECTOR_CLOB column populated by populate_vectors.py for semantic search.

Setup & running

Requires an Oracle Autonomous Database instance with OCI Generative AI access.

pip install -r requirements.txt
cp .env.example .env   # fill in your own credentials — never commit this file

# one-time: pre-compute product embeddings for semantic search
python populate_vectors.py

streamlit run app.py

Tech stack

Python · Streamlit · Oracle Cloud (Autonomous Database + Generative AI) · oracledb · oci SDK · pandas

Possible next steps

• Add automated tests for the SQL validator
• Cache repeated similarity searches instead of recomputing per query
• Extend the aggregated charts view with additional breakdowns