cassetteql 0.1.1

S3-native cassette substrate for typed knowledge graphs: Dempster-Shafer reasoner, sheaf GNN, Kan-based schema migration. No GPU, 17MB SPLADE bundled.

cassetteql

S3-native knowledge graphs with calibrated reasoning. Immutable cassette files, split Parquet indexes, a Dempster–Shafer reasoner that tells you when it doesn't know, a trained sheaf GNN prior, and Kan-based schema migration. The bundled SPLADE-tiny model (17 MB) means no GPU, no model download, no API keys.

  ┌──────────────────────────────────┐
  │  ▓▓      .inf  cassette      ▓▓  │  one cassette =
  │  ●                            ●  │  one ingest batch
  │  ╲╲─────────────────────────╱╱   │
  │   header  ·  records  ·  footer  │
  │     │         │           │      │
  │     │         │           └─ JSON: offsets + stats
  │     │         └─ gzip frames (range-addressable)
  │     └─ schema_ref · created_at
  └──────────────────────────────────┘
      immutable · content-addressed · S3-native

pip install cassetteql

Imports as cognition (same pattern as pillow/PIL):

from cognition.cassette import InfonStore, Query, Analyst

One-minute start

from cognition.cassette import InfonStore, Query

store = InfonStore("./data/chips", schema_path="schema.json")

# Delta ingest — idempotent, reports coverage diagnostics for free.
result = store.ingest(documents)
print(result["report"].summary())

# Calibrated single-claim verdict.
v = store.ask(Query().where(subject="toyota",
                             predicate="invest",
                             object="solid_state"))
print(v.label, v.mass.supports, v.mass.theta)   # SUPPORTS 0.53 0.29

# Multi-hop MCTS with retraction-aware chain mass.
v = store.connect("toyota", "catl")

# One tree walk resolves connectivity to many targets.
vs = store.any_of("toyota", {"catl", "lg", "samsung", "sk_hynix"})

Swap the root URI for s3://bucket/prefix and the same code runs against S3:

pip install 'cassetteql[s3]'
store = InfonStore("s3://acme/chips", schema_path="schema.json")

What makes it different

Cassette substrate	Immutable content-addressed .inf files; split Parquet indexes per cassette; append-only manifest chain. Delta ingest never rewrites; time-travel snapshots cost one JSON read.
Calibrated verdicts	Every answer carries (supports, refutes, theta). On claims the corpus can't answer, θ → 1.0 and no range-gets are issued — the pruner short-circuits.
Sheaf GNN prior	140k-param encoder with per-relation-kind restriction maps, trained once on synthetic hypergraphs (no human labels). 99% on held-out, +94% over symbolic on reportive-edge anomalies.
Schema migration	SchemaFunctor(rename, merge, delete) rewrites cassettes under a new ontology via Kan pushforward. 60× faster than reingestion; old cassettes stay.
Strands Analyst	Nine tools exposed to any Strands agent: schema / ingest / report / ask / connect / any_of / findings. System prompt enforces source citation and honest NEI.

Optional extras

pip install 'cassetteql[s3]'       # S3 / GCS / Azure via fsspec
pip install 'cassetteql[agent]'    # Strands Analyst
pip install 'cassetteql[aws]'      # Lambda container deploy + S3
pip install 'cassetteql[all]'      # everything optional

Measured

Each row below is a reproducible probe — a standalone Python script that writes a temp store, runs the scenario, and asserts the result. Probes ship inside the source distribution.

	Symbolic only	With sheaf GNN
10-claim actor-to-actor eval	40%	100%
2000-sample synthgen held-out	88.5%	99.2%
Reportive-edge anomaly accuracy	6%	100%
Range-gets per MCTS query at 300 cassettes	20	1.4
Migration vs. reingest (10-infon store)	1245 ms	20 ms (62×)

Dependencies

Package	Purpose	Required
torch ≥ 2.0	Reasoner + GNN + SSL losses	yes
transformers ≥ 4.40	SPLADE tokenizer/model	yes
numpy ≥ 1.24	Linear algebra	yes
pyarrow ≥ 15	Cassette indexes	yes
fsspec ≥ 2024.1	Local + cloud paths	yes
s3fs ≥ 2024.1	S3 backend	via [s3]
strands-agents ≥ 1.0	Conversational Analyst	via [agent]
boto3 ≥ 1.28	Lambda deploy + ECR	via [aws]

17 MB SPLADE-tiny ships inside the wheel — one pip install, no follow-up download, no GPU.

License

Apache-2.0. The bundled SPLADE-tiny-msmarco model is also Apache-2.0.