Caner Olcay
All projects

ToPWR Assistant — Knowledge-graph RAG

Polish questions translated into read-only Cypher over a Neo4j knowledge graph, orchestrated with LangGraph and served through an MCP server.

Role
Machine Learning Engineer
Timeline
2025 – present
Status
Ongoing
Area
Backend / retrieval
ToPWR app (opens in a new tab)Source on GitHub (opens in a new tab)
Retrieval benchmarkPR #50 (opens in a new tab)

28 Polish questions about 7 real PWr faculties, each asked in four forms, against the same Neo4j graph.

100%of questions found the right faculty, up from 50%

Correct-entity hits out of 7 for each query form, before and after the fix
Query formBeforeAfter fix
Stored spelling
7/7
7/7
Polish diacritics
7/7
7/7
Lowercase
0/7
7/7
Lowercase + diacritics
0/7
7/7
Overall14/28 · 50%28/28 · 100%

One question from the run

Pokaż informacje o jednostce "wydział informatyki i telekomunikacji".

Before
WHERE n.title CONTAINS 'wydzial informatyki i telekomunikacji'→ 0 rows
After fix
WHERE toLower(n.title) CONTAINS toLower('wydzial informatyki i telekomunikacji')→ Wydzial Informatyki i Telekomunikacji

Overview

An assistant for the ToPWR university app, built with the Solvro student group. A React chat UI, a FastAPI layer, an MCP server, and a Neo4j knowledge graph let people ask questions in Polish and get answers backed by structured university data. I’ve had 15 pull requests merged into it so far.

merged PRs
15
Docker services
4
generated Cypher
Read-only

Problem

University information is scattered across departments, pages, and systems, which makes it slow to find — and people ask about it in inflected Polish.

Approach

A knowledge-graph retrieval workflow that routes the question, generates Cypher, runs it read-only against the graph, and grades the result before answering.

How it works

  1. 01

    Ask

    Question in Polish

  2. 02

    Route

    LangGraph + guardrails

  3. 03

    Generate

    LLM-written Cypher

  4. 04

    Query

    Neo4j, read-only, via MCP

  5. 05

    Grade

    Answer, or “no data”

React, FastAPI, a FastMCP server, and Neo4j run as a 4-service Docker stack. LangGraph orchestrates routing, Cypher generation, and grading with guardrails, while a Prefect flow processes documents and populates the knowledge graph.

Outcomes

  • Fixed inflected Polish (“semestrze”) never matching titles stored in the nominative, using a fuzzy full-text fallback and an LLM grader that says “no data” instead of guessing.
  • Locked generated Cypher to read-only with query validation and Neo4j READ mode, and capped rows in code rather than in the prompt.
  • Built the Prefect ETL that turns PDFs into graph nodes via OCR and LLM-written Cypher, in parallel and idempotently.

Challenges

  • Polish inflection: the words people type rarely match how titles are stored.
  • Letting an LLM write database queries without letting it write to the database.
  • Saying “no data” when the graph doesn’t know, instead of guessing.

Stack

Python · FastAPI · LangGraph · LangChain · FastMCP · Neo4j · Prefect · Docker