res-sum

A Python package leveraging LLMs for research evidence synthesis

res-sum takes a folder of PDF research papers and produces structured summaries of each one using Large Language Models. It extracts text, builds a knowledge graph of entities and relationships across your papers, and uses hybrid retrieval (vector search + graph traversal) to produce contextually grounded summaries.

Built with ecology in mind, but works for any scientific field.

Features

• Batch-summarize PDFs — point it at a folder, get a structured summary for each paper
• Knowledge graph — extracts entities and relationships from your papers using LLMs, stored as a queryable NetworkX graph
• Hybrid retrieval (GraphRAG) — combines vector similarity search (ChromaDB) with knowledge graph traversal
• Domain-aware prompting — ecology-specific Chain-of-Thought prompts; custom domains via YAML
• Multiple LLM providers — Ollama (local, free, default), Ollama Cloud, Groq, OpenAI, Anthropic
• Multiple output formats — DOCX, JSON, CSV
• Persistent storage — vector store + knowledge graph persist to disk; incremental ingestion for new papers

Installation

pip install res-sum

For additional LLM providers:

pip install res-sum[openai]        # OpenAI (GPT-4o)
pip install res-sum[anthropic]     # Anthropic (Claude)
pip install res-sum[ollama-cloud]  # Ollama Cloud API
pip install res-sum[all-providers] # All of the above

Default setup (Ollama — free, local, no API key)

If you have Ollama installed locally, res-sum works out of the box with no API key:

ollama pull llama3.2

That's it.

Quick start

Python API

from res_sum import ResSum

# Initialize (defaults: Ollama local, ecology domain)
rs = ResSum(
    llm_provider="ollama",       # or "ollama_cloud", "groq", "openai", "anthropic"
    domain="ecology",            # or "general", or path to custom YAML
)

# Ingest papers — extracts text, builds vector store + knowledge graph
rs.ingest_papers("./pdf_folder/")

# Summarize across all papers
summary = rs.summarize("What are the key findings on pollinator decline?")

# Or batch-summarize: one summary per paper, saved to disk
rs.summarize_papers(
    pdf_directory="./pdf_folder/",
    output_directory="./summaries/",
    output_format="docx",        # or "json", "csv"
)

Command line

# Batch summarize with Ollama (default)
res-sum summarize \
    --pdf_directory ./papers/ \
    --output_directory ./summaries/ \
    --domain ecology

# Use Groq instead (requires API key)
res-sum summarize \
    --pdf_directory ./papers/ \
    --output_directory ./summaries/ \
    --provider groq \
    --api_key $GROQ_API_KEY

# See available providers, models, and domains
res-sum info

LLM providers

Provider	API key needed	Rate limits	How to use
Ollama (default)	No	None (runs locally)	Install Ollama, pull a model
Ollama Cloud	Yes (OLLAMA_API_KEY)	Based on plan	--provider ollama_cloud
Groq	Yes (GROQ_API_KEY)	Free tier available	--provider groq
OpenAI	Yes (OPENAI_API_KEY)	Pay-per-use	--provider openai
Anthropic	Yes (ANTHROPIC_API_KEY)	Pay-per-use	--provider anthropic

API keys can be passed directly or set as environment variables. They are never stored by the package.

Setting up API keys

Option 1 — Environment variables (recommended):

# Add to your ~/.zshrc or ~/.bashrc
export OLLAMA_API_KEY="your-key-here"    # for Ollama Cloud
export GROQ_API_KEY="your-key-here"      # for Groq
export OPENAI_API_KEY="your-key-here"    # for OpenAI
export ANTHROPIC_API_KEY="your-key-here" # for Anthropic

Then just specify the provider — the key is picked up automatically:

rs = ResSum(llm_provider="ollama_cloud")

Option 2 — Pass directly:

rs = ResSum(
    llm_provider="ollama_cloud",
    api_key="your-ollama-cloud-key-here",
)

To get an Ollama Cloud API key, go to ollama.com/settings/keys.

Domain configurations

res-sum ships with two built-in domains:

• ecology (default) — entity types: Species, Location, Method, Metric, Concept, Temporal. Includes ecology-specific section headers (Study Area, Field Methods, Statistical Analysis, etc.) and a 6-step Chain-of-Thought prompt.
• general — broader entity types for any scientific field.

You can define your own domain with a YAML file:

# my_domain.yaml
name: biomedical
entity_types:
  - name: DRUG
    description: "Pharmaceutical compounds or treatments"
    examples: ["metformin", "aspirin"]
  - name: DISEASE
    description: "Medical conditions"
    examples: ["diabetes", "cancer"]
relationship_types:
  - TREATS
  - CAUSES
  - ASSOCIATED_WITH

rs = ResSum(domain="./my_domain.yaml")

Retrieval modes

Mode	What it does	Best for
hybrid (default)	Vector search + graph expansion + community context, re-ranked	General summarization
local	ChromaDB vector search only	Specific factual queries
graph	Graph traversal + vector lookup	Relational queries
global	Community-level summaries + vector search	Thematic synthesis across many papers

summary = rs.summarize("...", mode="hybrid")  # or "local", "graph", "global"

Explore your knowledge base

After ingesting papers, open an interactive dashboard to visualize and inspect everything:

rs.explore()  # opens in your browser

Or from the command line:

res-sum explore --data_dir ./knowledge_base

The dashboard has four tabs:

• Overview — papers ingested, chunk counts, entity type breakdown, graph stats
• Knowledge Graph — interactive graph visualization. Nodes colored by entity type, sized by connections. Click to see relationships, filter by type, search by name.
• Vector Store — browse all text chunks by paper. See which section each chunk came from, expand to read full text.
• Communities — entity clusters detected by the Leiden algorithm, with LLM-generated summaries explaining what connects each group.

It's a single HTML file — works offline, shareable with collaborators.

Programmatic access

# Query an entity
rs.query_graph("Canis lupus")

# Most connected entities
rs.get_central_entities(top_k=10)

# Community structure
rs.get_communities()

# Access the NetworkX graph directly
graph = rs.knowledge_graph.graph

The graph is saved as GraphML and can be imported into Neo4j or any graph visualization tool.

How it works

PDF files
  → Text extraction (pymupdf4llm — handles multi-column, tables)
  → Section detection (ecology-aware regex + Markdown headers)
  → Chunking (RecursiveCharacterTextSplitter)
  → ChromaDB (embed + store chunks)
  → LLM entity/relationship extraction → NetworkX knowledge graph
  → Community detection (Leiden/Louvain)
  → Hybrid retrieval (vector + graph + community)
  → LLM summarization (Chain-of-Thought prompting)
  → Output (DOCX / JSON / CSV)

All data persists to a data_dir/ folder. Adding new papers only processes what's new.

Requirements

• Python >= 3.9
• Ollama installed locally (for default provider), or an API key for another provider

Contributing

Issues and pull requests are welcome on GitHub.

License

MIT