GraphSyndesmos

Ask who knows what in a GitHub repo, and see the graph path behind every answer.

Built for Graph Hacks (WeMakeDevs x FalkorDB), Track 3: Company Brain.

The problem

Questions about a codebase are about connections: who wrote the pull requests that touched this file, which files change most often, who should review a change here. A plain LLM sees isolated text and guesses. Here the data lives in a graph, so the agent follows real relationships and can show its work.

How it works

question -> local LLM (Ollama) picks a tool -> Cypher query -> FalkorDB -> answer + query shown

Graph model (repo graph)

• Nodes: Person {login, commits}, PR {number, title, state}, Issue {number, title, state}, File {path}
• Edges: (Person)-[:AUTHORED]->(PR | Issue), (PR)-[:TOUCHES]->(File), (PR)-[:FIXES]->(Issue)

Agent tools (each one is a Cypher query over the graph)

Tool	Answers
top_contributors	Who authored the most PRs?
hot_files	Which files change most often?
module_insights	Which files match a path, with their PRs and contributors?
who_knows_module	Who has authored the most PRs touching a module?
prs_by_person	What has this person worked on?

FalkorDB is the core of the product: every answer comes from a graph traversal, and the Cypher is printed with each response so the reasoning can be checked.

Components

• fetch.py: downloads contributors, issues, PRs and PR files from the GitHub API (resumable)
• load.py: builds the graph in FalkorDB
• agent.py: the CLI agent (Ollama + graph tools)
• api.py: FastAPI backend (/health, /api/module?name=...)
• test_api.py: API tests

Setup

Requirements: Docker, Python 3.10+, Ollama.

# 1. FalkorDB (browser UI at http://localhost:3000)
docker run -d --name falkordb -p 6379:6379 -p 3000:3000 \
  -v falkordb_data:/var/lib/falkordb/data falkordb/falkordb:latest

# 2. Python environment
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# 3. Local LLM (free)
ollama pull llama3.2:3b

# 4. Get data and build the graph
export GH_TOKEN=your_github_token      # read access to public repos is enough
python fetch.py
python load.py

# 5. Ask questions
python agent.py

Optional: OLLAMA_MODEL=qwen2.5:7b python agent.py for a stronger model. API: uvicorn api:app --reload, tests: pytest.

Example

Ask: Who knows the most about anything related to graph?

[graph query] who_knows_module({'name': 'graph'})
MATCH (p:Person)-[:AUTHORED]->(pr:PR)-[:TOUCHES]->(f:File)
WHERE toLower(f.path) CONTAINS toLower($module)
RETURN p.login AS person, count(DISTINCT pr) AS prs
ORDER BY prs DESC LIMIT 5

Answer: AviAvni, swilly22, GomezGab, aniruddhaadak80, and alblancc know the graph-related files best.

Limitations and next steps

• Uses the first 100 PRs' file lists to keep API usage small, so file-level rankings are a sample, not the full history.
• Small local models sometimes choose the wrong tool; a larger model helps.
• Tests (pytest) need FalkorDB running with the repo graph loaded via load.py.
• Next: store questions and answers in the graph so the agent remembers across sessions (Track 2), and add review and issue-ownership tools.

Demo video

VIDEO_LINK_HERE