harshaAI / Retrieval application
RAG Agent
A Streamlit document assistant using recursive text splitting, MiniLM embeddings, persistent Chroma cosine search, document selection, and optional web-search augmentation.
View repository01 / COMPONENTS & DATA FLOW
Inside the system.
- 01Files → text
- 02Chunks → MiniLM
- 03Chroma retrieval
- 04Guardrails + answer
Multi-format ingestion
DocumentProcessor reads PDFs, Word documents, spreadsheets, JSON, CSV, Markdown, and text. Its default splitter uses 1,000-character chunks with 200-character overlap and preserves metadata.
Persistent semantic index
VectorStore embeds chunks with all-MiniLM-L6-v2 and stores them in a Chroma PersistentClient collection configured for cosine distance. Ingestion validates chunks, checks generated IDs, and adds records in batches.
Inspectable answer path
RAGPipeline checks the query, retrieves candidates, applies selected-document filtering, optionally searches the web, and builds an answer with separate document and web source collections.
02 / IMPLEMENTATION DECISIONS
Why it works this way.
Evidence and answer travel together
The result includes context quality, source records, guardrail warnings, and per-request metrics such as processing time and retrieved source counts. Those fields let the UI expose the route to an answer.
Selection filtering has a limit
The pipeline retrieves twice the requested candidate count and then filters by selected filenames. A highly selective request can still return fewer passages than requested.
03 / SCOPE & TRADEOFFS
The implementation boundary.
A public application implementation, not evidence of measured production throughput. Filename selection is a retrieval preference and should not be described as tenant authorization.
04 / READ THE IMPLEMENTATION