Releases·Tool Release
Dev releases 37M-param non-autoregressive graph extractor for LlamaIndex pipelines, ~0.013ms per sentence
A solo developer built a ~37M-parameter non-autoregressive decoder that extracts knowledge graphs from sentence embeddings in a single forward pass, as a LlamaIndex connector for GraphRAG ingestion. It avoids LLM-based extraction's token costs, latency, and malformed JSON retries; since no text generation occurs, broken output is structurally impossible.
1 engagement1 sourceTue, Sep 1, 2026, 05:46 PM
AI Overview
- Takes SONAR (Meta's multilingual embedding space) sentence embeddings and outputs entities plus typed relations directly, no text generation anywhere
- Motivated by GraphRAG ingestion costs at volume: per-token LLM extraction pricing, slowness, and retry loops for broken JSON
- Ships as a LlamaIndex connector class; the dev claims three lines to production
- Solo developer project shared on Reddit on 2026-09-01; no benchmark comparisons against LLM extraction quality are included in the post
Related
Social·Community
Dev distills Claude Sonnet extraction into two 287M GLiNER encoders, hits 0.90 F1 vs teacher's 0.935 at 2.3 docs/s
A practitioner distills LLM-based extraction of ~5M court decisions into two 287M encoders: a fine-tuned GLiNER span marker and a multiple-choice model for entity grouping, kinds, and actions. On 30 held-out decisions the student pipeline reaches 0.901 mention F1 versus a 0.935 LLM self-consistency ceiling, running ~2.3 decisions/second on one RTX 5090. They ask the community for debugging help before processing the full corpus.
A practitioner distills LLM-based extraction of ~5M court decisions into two 287M encoders: a fine-tuned GLiNER span marker and a multiple-choice model for entity grouping, kinds, and actions. On 30 held-out decisions the student pipeline reaches 0.901 mention F1 versus a 0.935 LLM self-consistency ceiling, running ~2.3 decisions/second on one RTX 5090. They ask the community for debugging help before processing the full corpus.
6 engagement1 sourceredditSun, Oct 4, 2026, 01:58 PM
Releases·Tool Release
ISONGraph serialization format improves GraphRAG multi-hop accuracy from 40% to 80%
A developer benchmarked 10 graph serialization formats for feeding knowledge graphs to LLMs, finding that format choice alone swings multi-hop accuracy from 40% to 80% and token cost by ~70%. JSON was among the worst performers. The author built ISONGraph, a property-graph format optimized for LLM comprehension, released under MIT license in 6 languages.
A developer benchmarked 10 graph serialization formats for feeding knowledge graphs to LLMs, finding that format choice alone swings multi-hop accuracy from 40% to 80% and token cost by ~70%. JSON was among the worst performers. The author built ISONGraph, a property-graph format optimized for LLM comprehension, released under MIT license in 6 languages.
5 engagement2 sourcesredditMon, Jul 27, 2026, 06:44 AM
Products·launched
Graph serialization format benchmark for GraphRAG chains
This project benchmarks 10 graph serialization formats (JSON, GraphML, RDF variants, edge lists, etc.) for GraphRAG chains, finding that format choice can swing multi-hop accuracy from ~40% to ~80% and that json.dumps() wastes ~70% of token budget on syntax. It provides actionable guidance for tuning the overlooked serialization step.
This project benchmarks 10 graph serialization formats (JSON, GraphML, RDF variants, edge lists, etc.) for GraphRAG chains, finding that format choice can swing multi-hop accuracy from ~40% to ~80% and that json.dumps() wastes ~70% of token budget on syntax. It provides actionable guidance for tuning the overlooked serialization step.
1 engagement1 sourceredditMon, Jul 27, 2026, 02:34 AM
Developments·Paper
RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM
RAGU is an open-source modular GraphRAG engine that improves knowledge graph construction by separating extraction from consolidation. It uses two-stage typed extraction, DBSCAN deduplication, LLM summarization, and Leiden community detection, and is motivated by a compact extractor LLM.
RAGU is an open-source modular GraphRAG engine that improves knowledge graph construction by separating extraction from consolidation. It uses two-stage typed extraction, DBSCAN deduplication, LLM summarization, and Leiden community detection, and is motivated by a compact extractor LLM.
7 engagement2 sourcesarxiv, redditMon, Jul 13, 2026, 03:20 PM
Products·launched
Hillock: FOSS neuro-symbolic memory engine for local LLMs
Hillock is an open-source (AGPL-3.0) memory engine that ingests documents into a SQLite knowledge graph using a non-generative CUDA tensor pipeline (TALON) instead of LLM calls, enabling sub-second ingestion with low VRAM. It uses vector symbolic architecture for fast query gating and coreference resolution, calling local LLMs (via Ollama) only for final response rendering. It solves the bottleneck of slow, VRAM-heavy generative extraction for local LLM workflows.
Hillock is an open-source (AGPL-3.0) memory engine that ingests documents into a SQLite knowledge graph using a non-generative CUDA tensor pipeline (TALON) instead of LLM calls, enabling sub-second ingestion with low VRAM. It uses vector symbolic architecture for fast query gating and coreference resolution, calling local LLMs (via Ollama) only for final response rendering. It solves the bottleneck of slow, VRAM-heavy generative extraction for local LLM workflows.
22014 engagement4 sourcesredditTue, Aug 11, 2026, 07:05 PM