RAG pipeline architecture, smaller model, LangChain 1.x

Modular Pipeline Architecture

• Refactored into separate modules: retrieve.py (similarity search), generate.py (LLM interaction), pipeline.py (orchestration)
• Clear separation of concerns allows swapping components independently

Model Change

• Switched from Mistral 7B to TinyLlama 1.1B for faster inference
• Reduced RAM requirements from 8 GB to 2 GB (tradeoff: lower answer quality)

Other improvements

• Handle empty retrieval results gracefully (skip LLM when no documents found)
• Cross-platform path handling using pathlib
• Various Python best practices and style fixes

LangChain 1.x Upgrade

• Migrated from LangChain 0.3.x to 1.x
• Updated all imports for the new package structure

Components

langchain == 1.2.3
langchain-classic == 1.0.1
langchain-community == 0.4.1
langchain-text-splitters == 1.1.0
langchain-huggingface == 1.0.0
langchain-chroma == 1.0.0
chromadb == 1.4.0
unstructured[local-inference] ~= 0.18.27
markdown ~= 3.10
sentence_transformers == 5.2.0
gpt4all == 2.8.2
streamlit ~= 1.52.2

LangChain v0.3, Sentence Transformers v3, GPT4All v2.8

Tagging a stable release using these components:

langchain == 0.3.13
langchain-community == 0.3.13
langchain-huggingface == 0.1.2
langchain-chroma == 0.1.4
chromadb == 0.5.23 # Vector database - embedding persistence
unstructured[local-inference] ~= 0.16.11
markdown ~= 3.7.0
sentence_transformers == 3.3.1
gpt4all == 2.8.2

LangChain 0.1, Sentence Transformers v2, GPT4All v2.5

Tagging a stable release using these components:

langchain == 0.1.16
langchain-community == 0.0.34
chromadb == 0.4.24 # Vector database - embedding persistence
unstructured[local-inference] ~= 0.12.5
markdown ~= 3.0.0
sentence_transformers == 2.5.1
gpt4all == 2.5.1