DeliVali/cogito-estella

Cogito Estella — Latent Graph Engine (v0.7.0)

Non-autoregressive decoder heads that turn SONAR (Meta) sentence embeddings directly into knowledge-graph triples — no token decoding, structurally valid by construction, 0.013 ms per concept on an RTX 5070.

Code, tests, and quickstart: https://github.com/DeliVali/cogito-estella

Tool-call demo (F1 = 1.000)

Code demo (F1 = 0.781)

Checkpoints

File	Modality	Triple F1	Params
cogito-toolcalls-graphdecoder.pt	tool-calls / API control	1.000	24.3M
cogito-code-lora-adapters.pt	Python code (LoRA + decoder)	0.781	20M + 26M
cogito-prose-candidates-{ft,cal,base,s2,s3}.pt	entity-conditioned prose (ensemble of 5)	0.827	5 × 37M
cogito-prose-openvocab{,-s4,-s5}.pt + cogito-prose-cascade-fallback.pt	open-vocab prose stack	0.6514	3 × 45M + 37M
vocab-prose.json	20k entities · 60 relations	—	—

Held-out by combination (no duplicate leakage); prose numbers validated on virgin slices of a 119,911-sample pool never used for model, threshold, or ensemble selection.

Usage

import json, torch
from cogito_estella.model.candidate_decoder import (
    CandidateDecoderConfig, CandidateGraphDecoder, decode_triples_coo)
from sonar.inference_pipelines.text import TextToEmbeddingModelPipeline

vocab = json.load(open("vocab-prose.json"))
rels = sorted(vocab["rel2id"], key=vocab["rel2id"].get)
dec = CandidateGraphDecoder(CandidateDecoderConfig(n_relations=len(rels))).cuda()
dec.load_state_dict(torch.load("cogito-prose-candidates-ft.pt")["dec"])
dec.eval()

pipe = TextToEmbeddingModelPipeline(encoder="text_sonar_basic_encoder",
                                    tokenizer="text_sonar_basic_encoder",
                                    device=torch.device("cuda"))
emb = pipe.predict(["The committee approved the new budget."], source_lang="eng_Latn")
cands = ["committee", "budget", "hospital"]                    # noun scan + graph nodes
ids = torch.tensor([[vocab["ent2id"][c] for c in cands]]).cuda()
mask = torch.ones(1, len(cands), dtype=torch.bool).cuda()
out = dec(emb.float().cuda(), ids, mask)
for i, r, j in decode_triples_coo(out["exist_logits"].cpu(), out["adj_logits"].cpu(),
                                  mask.cpu(), threshold=0.15, force_top1=True)[0]:
    print(cands[i], rels[r], cands[j])       # committee support budget

Intended use & limitations

Structured-knowledge extraction for GraphRAG ingestion, agent long-term memory, tool-call observability, and codebase mapping. Entities are selected from caller-supplied candidates (entity-conditioned head) or a 20k-lemma vocabulary (open-vocab head); relations come from a 60-verb vocabulary — content outside those spaces is not captured. Prose oracle is dependency-parse SVO; graphs on free prose inherit its noise profile. The SONAR encoder is required at inference and is distributed separately by Meta under CC-BY-NC 4.0; the LoRA adapters in cogito-code-lora-adapters.pt modify that encoder and inherit its non-commercial terms. All from-scratch decoder heads in this repository are Apache-2.0.