csg
Résumé ↗
Menu
← All projects

Retrieval · Voice / Configurable prototype

Voice-Based RAG

Ask your documents. Hear the answer.

A configurable document question-answering pipeline with text or voice input, hybrid retrieval, reranking and speech output.

Python · Whisper · Chroma · Ollama · Gradio

View repository ↗

01 / THE PROBLEM

What needed solving.

Documents arrive in different formats, while people may prefer speaking to typing. A reusable retrieval system must normalize inputs, select relevant context and keep speech handling separate from retrieval and generation.

02 / FOLLOW THE WORKFLOW

See how it fits together.

A simplified example of the documented flow. Sample outputs are authored illustrations, not live model predictions.

Voice-Based RAGIllustrative workflow

Multi-format document ingestion

Handbook.pdf → searchable passages

Parse and chunk source documents.

Example query

“How do I submit an expense?”

Accept a typed or transcribed question.

Context selected from the documents

Dense + BM25 → ranked passages

Combine retrieval and reranking.

Illustrative pipeline, no microphone used

Retrieved context → answer → audio

Generate a response with optional speech.

03 / ENGINEERING

The decisions behind it.

Make stages replaceable

Parser, chunker, embedding, retrieval, model and speech modules have separate boundaries. Per-format chunking and YAML configuration let the pipeline adapt to the corpus.

Retrieve broadly, then rerank

Dense retrieval can combine with BM25, followed by a cross-encoder. Chroma is the default vector store; FAISS and Qdrant are configurable alternatives.

Keep voice optional

Whisper variants handle transcription, and configurable TTS engines provide output. Text mode and muted speech remain available. Ollama supports a local runtime; Groq is an optional hosted provider.

04 / EVIDENCE & LIMITS

Claims with context.

3Documented vector-store options
Text + voiceInput and response workflows

The README documents end-to-end orchestration through PipelineOrchestrator and Gradio, with multi-format parsers, configurable chunkers, hybrid retrieval and reranking. It reports implementation and tests, without a fixed-corpus Recall@k or grounded-answer benchmark.

Current limits & next validation

Retrieval quality depends on corpus preparation, chunking and model choice. A useful next evaluation compares dense, hybrid and reranked retrieval on the same fixed questions and sources, with answer grounding and end-to-end latency measured separately.

05 / READ THE SOURCE

Go deeper.

Based on public repository documentation reviewed on 7 November 2026. Results are attributed to the project’s recorded evaluations; they were not rerun for this portfolio.

NEXT PROJECTReview Tag Generator

TOOLS BEHIND THE WORK

Explore the stack. Drag the ribbon or select a tool.