Building Full GenAI Applications

Extract clean text, page or section metadata, and document identifiers from PDF and plain-text files.

Split extracted text into retrieval chunks with configurable size, overlap, and stable source identifiers.

Generate chunk embeddings with a local embedding model and use cosine similarity to compare semantic relevance.

Store embeddings and chunk metadata in a persistent vector index, then reload the index in a new application process.

Retrieve the top relevant chunks for a user question and inspect their similarity scores.

Construct a grounded prompt that clearly delimits retrieved context as untrusted data and requires source identifiers in supported answers.

Implement a callable RAG pipeline that combines retrieval, the selected model provider, and citations validated against the retrieved chunks.

Apply an evidence check that returns an explicit insufficient-evidence response for unsupported questions while preserving valid answers.

Implement a FastAPI ingestion endpoint that accepts a PDF or text document and builds its vector index.

Implement a FastAPI question-answering endpoint that runs retrieval and grounded generation as a streamed response.

Build a Streamlit interface for document upload, model selection, and conversational questions.

Connect Streamlit to the FastAPI endpoints and display incremental answer text with validated source citations.

Maintain bounded conversation histories and separate document indexes for independent user sessions.

Handle provider timeouts, rate limits, unavailable local models, and malformed responses without crashing the application.

Enforce allowed file types, file-size ceilings, question and context length limits, and safe error messages at the API boundary.

Record structured logs for request latency, selected provider, retrieval results, model usage, and failures without exposing document text or secrets.

Create a small evaluation set containing answerable, ambiguous, and unanswerable questions for a sample document.

Measure whether retrieval returns the source passages needed to answer each evaluation question.

Assess generated answers for groundedness, citation correctness, and appropriate refusal when evidence is missing.

Compare Groq, Mistral, and Ollama results by answer quality, latency, privacy, and estimated cost.

Tune chunk size, overlap, retrieval count, evidence threshold, and prompt instructions in response to observed failures.

Write focused automated tests for chunking, retrieval, provider switching, evidence handling, and the question-answering API.

Rebuild the core retrieval-and-generation workflow with LangChain and map each abstraction to the corresponding raw-Python component.

Containerize and launch the FastAPI and Streamlit services with Docker Compose using environment-based hosted-provider configuration.