RAG-Structured-Data

Retreival-Augmented Generation (RAG) stack that allows a user to have a conversational interface over structured datasets such as Excel files, CSVs or JSONs -- without summarizing or losing granularity of the original data.

Features

• Ingest various types of data

• Support queries that require

  • Row specific retrieval
  • Column based filtering
  • Numeric Lookups
  • Range based queries

• Pre- and post-optimization steps to ensure the retrieval preserves exact data values and context.

• Handle numeric precision and unit consistency.

• Avoids hallucination from the model

• Uses FAISS for indexing and semantic search retrieval

• Evaluates among various indexing strategies to get the best one based on metrics like Precision@1 and MRR@k

• Hybrid reranker using Cosine similarity + Keyword overlap

• Evaluates whether a user query requires semantic or symbolic filters (uses Pandas for symbolic filtering) for improved performance

• Provides provenance for each query so you are sure that your response is from your data itself

• Uses interpolation to answer range based queries

• Converts the query into a deterministic plan so that the LLM cannot fabricate any values

• LLM based conversational interface

• Maintains the conversation history

• Streamlit UI for easy uploading and overall UX

---

Architecture

---

Demo Video

Pipeline Architecture

Ingestion and Canonicalization

• Normalizes column names

• Extracts units from headers but keeps a map for indexing later

• Enforces dtypes

• Adds a stable __row_id to enable provenance

• Persists a parquet copy for speed and a schema.json with column metadata and units

Indexing

• Convert rows to compact text representations

• Embed rows with SentenceTransformers

• Build a FAISS index and save it with a stable mapping to DataFrame row ids

• Reload the index and run top-k semantic search

Query Planner

• Parse simple numeric filters directly using regex

• Use the OpenAI API to produce a json plan which describes what the retriever should do:

  • filter
  • range
  • aggregate
  • plot
  • semantic search

Retriever and Executor

• Runs numeric filters precisely with pandas

• Handles range queries with linear interpolation

• Runs semantic search via FAISS index and returns exact rows with provenance

• Provides an optional plot spec that can later be rendered

Answerer and Guardrails

• Retrieves the result from earlier module

• Formats the answer without using an LLM that explains the retrieved result

• Answers using an optional LLM rephraser that is hard-guardrailed to only use provided rows/values

• Refuses when there is no supporting data

Retrieval optimizations & comparisons

• Build multiple semantic indexes using the following strategies

  • Only rows
  • Rows + Headers
  • Cell Facts (Dense)

• Hybrid reranker: cosine score + keyword overlap

• Evaluates on a small eval.jsonl that contains sample queries with expected ground truths

• Reports Precision@1, MRR@K, and median/mean latency per strategy