Felix

│

AI-powered continuity assistant for screenwriters working on complex, multi-era narratives.

status: paused github: repo

Why

A friend writes screenplays — complex thrillers with interleaved timelines, dozens of characters, and continuity traps everywhere. He was tracking everything in spreadsheets. I figured a graph database + LLM could do better: parse raw scene text, build a knowledge graph, and let the writer query it in natural language.

What it does

Drop raw text in, Felix handles the rest:

▸Extract — LLM agents pull entities, relationships and events out of raw text
▸Graph — Fuzzy-merges everything into Neo4j. The schema isn't hardcoded: a "profile" defines the entity types and relation vocabulary of the domain
▸Check — Bilocalization and timeline contradictions ("he acts two scenes after his death") detected by reasoning over each entity's ordered event chronology
▸Query — Natural language questions, entity sheets generated straight from the graph

From 100% local to API (the pivot)

The project started with a hard constraint: everything had to run locally on a Mac M4, with Qwen2.5-7B via LM Studio. Entity extraction held up thanks to heavy guardrails — narrative reasoning on a 7B did not. Rather than keep fighting the hardware, I pivoted to the Mistral API (mistral-small), keeping the small-model-plus- targeted-tools philosophy. The guardrails survived the pivot: they turned out to be the real product.

The real challenge: caging the LLM

An LLM writing into a graph will invent entities, relationships, plot twists. The whole job is stopping it:

▸Typed relations, hard vocabulary — enforced at graph-write time: a relation type outside the vocabulary is rejected, not stored
▸Time-bounded states are events — death, prison, a job: ordered events, not properties. The consistency checker reasons over them deterministically
▸Eval-first — every engine behavior has a replayable eval built with apte: unit-level, integration probes, full end-to-end scenarios

Stack

▸FastAPI + Neo4j 5 (async) + ChromaDB
▸Pydantic AI agents + Mistral API (LM Studio still pluggable for local)
▸Nuxt 3 + Nuxt UI ("paper" front-end)
▸apte for the eval suite
▸Python 3.12, mypy strict, Ruff

What I learned

Graph modeling for narrative data is surprisingly tricky — when does a character mention become a node rather than a property? How do you merge entities across scenes written months apart?

A free-form relation vocabulary always drifts — validating it hard at write time killed an entire class of bugs at once. And you can't steer an LLM agent without an eval harness: here, every fix starts with a failing eval.

The wall

I hit a pretty hard wall between what I wanted to do and what the small models I was using could do. Extraction holds. Narrative reasoning, merging entities across scenes written months apart, restraint (writing nothing on a plain "hi"): no. A 7B invents, and mistral-small still invents. The guardrails limit the damage, they don't replace a model that understands the story.

Status

Paused since June 2026, resume point documented in the journal. The engine runs on a real screenplay and "screenplay" is just one profile among others (a "construction site" profile exists).

github.com/renaudcepre
FR