Cogito Estella: Latent Graph Engine (v0.10.0)

Non-autoregressive inference backend that decodes SONAR (Meta) semantic embeddings directly into knowledge graphs, bypassing token-based text decoding entirely. Mapped, trained, and validated end-to-end on a single RTX 5070 (12 GB).

Why

GraphRAG over millions of documents with autoregressive LLMs is budget-prohibitive: per-token decoding, per-token API pricing, and malformed-output retry loops. Cogito Estella emits the full adjacency structure in one dense forward pass — structurally valid by construction, at consumer-hardware cost.

Inference Architecture

The engine attaches compact non-autoregressive decoder heads (5.8M–45M parameters) to Meta's frozen multilingual latent space. Input: a concept embedding $e \in \mathbb{R}^{1024}$ (sentence segmented with SaT, projected by SONAR). No time dimension — a single pass emits the complete graph.

e ∈ ℝ^1024
  │  deep trunk (optional): 1–3 × [Linear → GELU → LayerNorm]
  ▼
N ∈ ℝ^{K × d}          # K node slots (fixed grid) or C candidates (elastic)
  ├─ existence   s ∈ ℝ^K            = σ(W_s · N)
  ├─ labels      ℓ ∈ ℝ^{K × V}      = W_ℓ · N          (open-vocab head)
  └─ adjacency   A ∈ ℝ^{R × K × K}  = N_i^T W_r N_j    (low-rank: W_r = U_r V_r^T)

Two production head families:

• GraphDecoder (open-vocab): fixed K-slot grid, V-way label softmax per slot, bilinear adjacency. Optional deep GELU/LayerNorm trunk (trunk_layers).
• CandidateGraphDecoder (entity-conditioned): caller supplies candidate entities (text scan and/or graph-memory nodes); candidates act as cross-attention queries over learned concept views via nn.Embedding lookup. Output space is restricted by construction — zero vocabulary leakage. Elastic COO adjacency sized to the candidate list; low-rank bilinear relations (rank 64); calibrated decode with a force-top1 recall floor.

Training objective — structural loss, no sampling, no teacher forcing: $\mathcal{L} = \mathrm{BCE}(\hat{s}, s) + \mathrm{CE}_{\text{masked}}(\hat{\ell}, \ell) + \mathrm{BCE}(\hat{A}, A)$.

Metrics & Experimental Benchmark Summary

Held-out evaluation, split by combination (an entity/relation combination never seen in training, eliminating duplicate leakage). Prose numbers are validated on virgin slices of a 119,911-sample held-out pool that took no part in any model, threshold, or ensemble selection; structured modalities use their own combination-held splits (tool-calls: 12,147 held-out concepts spanning 240 unseen combinations).

Oracles: exact-by-construction JSON (tool-calls), native Python ast parser (code), deterministic dependency-parse SVO (prose). Adjacency thresholding ships three strategies (fixed, otsu, noise_floor); the sparsity-prior noise_floor dominates variance-based Otsu on sparse graphs (8/8 vs 3/8 stress scenarios).

Core Latency & Compute Footprint

Measured on consumer silicon (RTX 5070, 12 GB, CUDA 12.8), batch 1024:

• Non-autoregressive forward pass: 0.013 ms per sentence concept (single CandidateGraphDecoder; 0.052 ms for the 24.3M open-vocab production config; ~0.06 ms for the 5-model prose ensemble). 269–1074× faster wall-clock than autoregressive text decoding of the same content.
• Context state: 336 KB per active prefill thread vs. 2.1 GB KV-cache on a standard 7B autoregressive model at 4k context — a 6,400× reduction.
• FLOP reduction: 36× against a matched char-level token baseline (dense structural blocks) up to 2,881× for exact string literals via the parallel char-grid mapping (digits/characters as nodes, no copy loop).
• Encode throughput (SaT + sanitization + SONAR): ~273 concepts/s; storage density ~2.19 KB/concept (fp16).

Production Integration Patterns

• GraphRAG ingestion pipelines — vector-to-knowledge-graph indexing without per-token commercial API calls; graphs land as COO triples ready for MERGE into any property-graph store, with per-edge provenance.
• Non-autoregressive tool dispatchers — raw text streams decoded into typed call structures in one pass; output is structurally valid by construction (no JSON repair, no syntax-breaking retries).
• Incremental AST repository tracking — hook file-save events to re-encode changed units and merge live call/import dependency sub-graphs into memory.
• Multilingual agent memory shards — chat history compressed into language-agnostic 1024-d concepts (SONAR, 200 languages) and decoded to a queryable graph; entity-conditioning accepts the agent's existing graph nodes as candidates.

Use Cases

• Agent long-term memory ("second brain") — every message, note, or document an agent touches becomes graph triples in a property store (Neo4j, SQLite); sessions end, knowledge persists and stays queryable.
• GraphRAG grounding — answer-time retrieval over structured facts instead of (or alongside) vector similarity; the LLM cites edges that exist, reducing hallucinated recall.
• Tool-call observability — agent traces decoded at F1 1.000 make "which tools touched service X" an exact graph query, at near-zero compute.
• Codebase intelligence — call/import dependency graphs over whole repositories at a cost where re-indexing on every save is affordable.
• Multilingual knowledge consolidation — SONAR's language-agnostic space lets documents in 200 languages land in one shared graph.
• Log & trace mining — high-volume streams (support tickets, incident reports, transcripts) structured continuously on one consumer GPU, where per-token LLM extraction is cost-prohibitive.
• Compliance & provenance KBs — every edge carries source metadata; audits answer "where does this fact come from" by construction.

Global Summarization (GraphSummarizer)

GraphRAG-style sensemaking where the expensive stage (per-chunk extraction) costs zero LLM tokens; the LLM only writes per-community summaries (~10 calls per book instead of ~600), with automatic entity-grounding rejection of unverifiable output:

from cogito_estella.graph_summary import GraphSummarizer

gs = GraphSummarizer(llm_fn=my_llm)          # any callable: prompt -> str
report = gs.summarize(triples, top_k=8)      # triples from the extractor
for c in report:
    print(c["accepted"], c["grounding"], c["top_entities"][:5], c["summary"])

Validated on a full novel: coherent thematic clusters at ~155x fewer LLM tokens than per-chunk extraction pipelines; low-cohesion clusters are gated out and summaries that mention entities absent from their cluster are rejected automatically.

Pretrained Weights & Quickstart

Champion checkpoints ship via GitHub Releases and Hugging Face (DeliVali/cogito-estella):

pip install "cogito-estella[sonar]"     # or: uv sync (from a clone)
# download cogito-prose-candidates-ft.pt + vocab-prose.json next to quickstart.py
python quickstart.py

quickstart.py encodes two raw sentences and prints their decoded triples in one non-autoregressive pass.

Exact literals (phones, hashes, IDs, precise amounts) never round-trip through the embedding: extract_with_literals detects them deterministically in the source text and literals_to_neo4j stores them verbatim with provenance — character-exact recovery by query, guaranteed by copying rather than decoding.

Weight licensing: all from-scratch decoder heads (GraphDecoder, CandidateGraphDecoder, trunks, ensembles) are Apache-2.0. The code-modality LoRA adapters modify Meta's SONAR encoder, whose weights are distributed under CC-BY-NC 4.0 — the adapted encoder inherits those non-commercial terms.

The Map of Literals (Open-Vocab Handling)

Continuous embeddings lose exact values (measured: MAE 36, 1.5% exact on held-out integers via linear probe). The ceiling breaks at the data level, not the model:

• Integers: digits-as-nodes + atomic spacing ("400" → "4 0 0"): 98.3% exact on unseen values.
• Short strings (≤4 chars): characters-as-nodes: 89.9% exact, open vocab.
• Long arbitrary strings: hybrid copy channel (the only autoregressive path).

Core Structure (src/cogito_estella/)

src/cogito_estella/
  model/graph_decoder.py       # non-AR GraphDecoder + deep trunk + decode strategies
  model/candidate_decoder.py   # entity-conditioned head: cross-attention, COO, force-top1
  model/transformer.py         # ConceptTransformer (RoPE, RMSNorm, SwiGLU)
  model/token_baseline.py      # matched-FLOP token baseline
  model/sonar_loss.py          # CE propagated through the frozen SONAR decoder
  model/train_graph.py         # resumable training loop
  sonar_codec.py               # SONAR encode/decode, OOM-resilient
  segmenter.py                 # SaT multilingual segmentation
  multilingual_factory.py      # canonicalized ingestion from open sources
  preprocess.py                # sanitization + atomic spacing (script-gated)
  graph_target.py              # target-graph construction
  graph_metrics.py             # Triple F1, GED proxy, tool-call F1
  compute.py                   # FLOP/bandwidth accounting, roofline

Training data, experiment scaffolding, and logs are untracked; the unit-test suite ships with the repository — 113 tests, fully reproducible offline:

uv sync
uv run pytest tests/

Requires Python 3.12; SONAR/SaT weights download on first use.

License

Distributed under the Apache License 2.0. See the LICENSE file for details on relational-patent protection and commercial use.