neuromind 0.1.0

Neuromind RAG research library: download PMC papers, index into ChromaDB, and query with multi-level RAG

neuromind

An umbrella Python library for biomedical RAG research.

Sub-packages

Package	Description
neuromind.alzheimers	Alzheimer's disease RAG pipeline

---

neuromind.alzheimers

Combines PubMed Central paper downloading, ChromaDB vector indexing, and a multi-level Retrieval-Augmented Generation (RAG) pipeline.

Built on:

• pubmed-stream — concurrent PMC article downloads via NCBI E-utilities
• ChromaDB — local persistent vector store
• Ollama — local LLM inference (128K context)

Inspired by the RAG_Comparison_Demo.ipynb notebook, which demonstrates six evidence-depth conditions from No RAG through Comprehensive RAG on Alzheimer's disease / gut-microbiome research questions.

---

Installation

# Core library
pip install .

# With Ollama Python client
pip install ".[llm]"

# With biomedical sentence-transformers embeddings
pip install ".[embeddings]"

# Everything
pip install ".[all]"

Requirements

• Python ≥ 3.8
• Runtime: requests, chromadb, pubmed-stream
• Optional: ollama (Python package), sentence-transformers
• LLM server: Ollama running locally

---

Quick Start

Python API

from neuromind.alzheimers import AlzheimersRAG, RAGMode
# or, via top-level re-exports:
from neuromind import AlzheimersRAG, RAGMode

rag = AlzheimersRAG(
    db_path="./neuromind_db",
    model="llama3.2",          # any model available in Ollama
)

# 1. Download papers from PubMed Central
rag.download("Alzheimer's disease gut microbiome", max_results=50)

# 2. Index into ChromaDB
stats = rag.index()
print(stats)  # IndexStats(indexed=48, chunks=1_024, ...)

# 3. Query with RAG (default: DETAILED — 15 papers, ~70K context tokens)
result = rag.query(
    "What is the evidence for gut microbiome dysbiosis in Alzheimer's disease?",
    mode="detailed",
)
print(result.text)
print(result.summary())   # [detailed] 612 words | 15 sources | 68,402 ctx tokens | 42.3s

CLI

# Download papers
alzheimers download "Alzheimer's disease gut microbiome" --max-results 100

# Download the full built-in AD corpus (15 topics)
alzheimers download --corpus --max-results 100

# Index into ChromaDB
alzheimers index ./publications --db-path ./neuromind_db

# Query
alzheimers query "What is the role of tau in Alzheimer's?" --mode detailed

# Compare all RAG modes on the same question
alzheimers compare "What therapeutic strategies target the gut-brain axis in AD?"

# python -m neuromind.alzheimers also works
python -m neuromind.alzheimers query "..." --model llama3.2

---

RAG Modes

Mode	Papers	~Context tokens	Use case
no_rag	0	0	Baseline / hallucination check
retrieval_only	5	0	Inspect raw evidence
basic	5	~20 K	Fast, grounded facts
standard	10	~40 K	Daily clinical queries
detailed ⭐	15	~70 K	Recommended default
comprehensive	25	~100 K	Research-grade depth

---

Architecture

User Query
    │
    ├─► [No RAG]          LLM only (training knowledge)
    │
    └─► [RAG modes]       pubmed-stream → ChromaDB → LLM
                               │               │
                         NCBI E-utilities   MedEmbed /
                         (ESearch + EFetch) MiniLM-L6-v2

Data flow for RAG modes:

query  →  Retriever.retrieve(top_k)
              │
              ▼
        ChromaDB (cosine similarity)
              │
              ▼
        build_context (token budget)
              │
              ▼
        OllamaLLM.generate(prompt)
              │
              ▼
        RAGResult (.text, .sources, .summary())

---

Python API Reference

AlzheimersRAG

rag = AlzheimersRAG(
    db_path="./neuromind_db",   # ChromaDB location
    collection_name="publications",
    model="llama3.2",            # Ollama model tag
    ollama_host="http://localhost:11434",
    num_ctx=131_072,             # LLM context window
    embedding_fn=None,           # custom ChromaDB embedding function
)

Method	Description
rag.download(keyword, max_results, ...)	Download PMC articles
rag.download_corpus(topics, max_results_per_topic, ...)	Download full AD corpus
rag.index(papers_dir, chunk_size, chunk_overlap)	Index into ChromaDB
rag.query(question, mode, max_response_tokens, temperature)	RAG query
rag.compare_modes(question, modes, ...)	Run all modes, return dict

Low-level building blocks

from neuromind.alzheimers import download_papers, download_ad_corpus
from neuromind.alzheimers import index_directory
from neuromind.alzheimers.retriever import Retriever
from neuromind.alzheimers.llm import OllamaLLM
from neuromind.alzheimers.rag import RAGPipeline
from neuromind.alzheimers.types import RAGMode, RAGResult, IndexStats

---

Custom Embeddings (e.g. MedEmbed)

To use the biomedical MedEmbed embeddings from the original notebook:

from medembed_embedder import LangchainMedEmbedEmbeddings
import chromadb

# Wrap MedEmbed as a ChromaDB embedding function
class MedEmbedFn:
    def __init__(self):
        self._model = LangchainMedEmbedEmbeddings(device="cpu")
    def __call__(self, input):          # ChromaDB EmbeddingFunction protocol
        return self._model.embed_documents(input)

from neuromind.alzheimers import AlzheimersRAG
rag = AlzheimersRAG(db_path="./neuromind_db", embedding_fn=MedEmbedFn())

---

NCBI API Key

Vector Database Path

Point AlzheimersRAG at your ChromaDB directory via the NEUROMIND_DB_PATH environment variable instead of passing db_path every time:

export NEUROMIND_DB_PATH="/path/to/your/knowledge_base"

from neuromind.alzheimers import AlzheimersRAG
rag = AlzheimersRAG()          # picks up NEUROMIND_DB_PATH automatically
rag = AlzheimersRAG(db_path="/explicit/override")  # or pass it directly

If the variable is not set the default is ./neuromind_db.

---

NCBI API Key

An NCBI API key raises the rate limit from 3 → 10 requests/second, making downloads ~3× faster.

export NCBI_API_KEY="your_key_here"
export NCBI_EMAIL="[e-mail]"

Get a free key at https://www.ncbi.nlm.nih.gov/account/.

---

Project Structure

neuromind-project/
├── neuromind/
│   ├── __init__.py           # umbrella re-exports
│   └── alzheimers/
│       ├── __init__.py       # AlzheimersRAG façade + public API
│       ├── __main__.py       # python -m neuromind.alzheimers
│       ├── cli.py            # CLI (download / index / query / compare)
│       ├── downloader.py     # pubmed-stream wrapper + AD_SEARCH_TERMS
│       ├── indexer.py        # text chunking + ChromaDB upsert
│       ├── retriever.py      # ChromaDB semantic retrieval
│       ├── llm.py            # Ollama Python client / HTTP fallback
│       ├── rag.py            # RAGPipeline (all 6 modes)
│       └── types.py          # RAGMode, RAGResult, IndexStats, …
├── examples/
│   └── basic_usage.py
├── RAG_Comparison_Demo.ipynb
├── pyproject.toml
└── README.md

---

License

MIT — see LICENSE.