Source-linked AI summary
Noesis: Bidirectional Graph-RAG with Adaptive Parallelism and Cross-Knowledge-Base Semantic Discovery
Nicola Cogotti
TL;DR
Existing Graph-RAG systems struggle with long-document structure, adaptive ingestion, and automatic routing across specialized knowledge bases. Noesis addresses these gaps with a decoupled architecture and four algorithms, surpassing GraphRAG by +27.8 EM on HotpotQA while achieving 90% source-verified precision on long-range causal edges.
Problem
Existing Graph-RAG systems lack robust long-document structure, adaptive document-level parallelism, and automatic cross-knowledge-base routing with runtime discovery.
Method
Noesis uses a decoupled architecture with bidirectional graph traversal, adaptive AIMD concurrency control, selective MoE quantization, and cross-knowledge-base semantic routing.
Results
+27.8 EM over GraphRAG on HotpotQA, while source-text verification achieves 90% precision on long-range causal edges from a 193-page document.
Takeaways & Limitations
Mesh demonstrates that architecture can enable emergent multi-hop cross-domain discovery for a 2.3B-parameter model that cannot perform it without Mesh.
Takeaways & Limitations
The Mesh evaluation demonstrates its mechanism on one representative query across three domain-specific knowledge bases, leaving broader query and configuration coverage uncharacterized.
Abstract
from arXiv · showhide
Retrieval-Augmented Generation over knowledge graphs (Graph-RAG) has emerged as a powerful paradigm for grounding large language models in domain-specific corpora. However, existing systems face persistent limitations: (1) static chunking fragments long documents, losing cross-section semantic connections; (2) ingestion pipelines do not scale adaptively; and (3) multi-domain deployments require either a monolithic knowledge base that dilutes retrieval precision or manual user routing. We present Noesis, a decoupled Graph-RAG architecture addressing these limitations through four algorithms: (a) Bidirectional Graph Traversal with a Graph-Feedback Context Resolver simulating human reading with degrading memory; (b) an AIMD Concurrency Controller adapted from TCP congestion control, achieving 23x speedup with zero OOM events; (c) Moesis, domain-aware selective quantization for MoE models achieving 6.3x speedup on 12 GB consumer GPUs; and (d) Mesh, cross-KB semantic routing with runtime structural discovery enabling small on-premises models to perform multi-hop cross-domain reasoning. On HotpotQA (1,000 questions), Noesis achieves 59.5 EM / 74.7 F1, surpassing GraphRAG by +27.8 EM while using a 35B on-premises model for graph construction rather than GPT-4o. Source text verification on a 193-page document confirms 90% precision on long-range causal edges inaccessible to chunk-independent extraction.
1 Introduction
The introduction identifies three production bottlenecks in Graph-RAG—semantic fragmentation, rigid parallelism, and cross-domain isolation—and presents Noesis as a system addressing them with four specialized contributions.
- Graph-RAG combines knowledge graphs with LLMs to support verifiable, multi-hop reasoning over complex corpora.Knowledge graphs provide structured domain expertise that grounds LLM responses in facts.
- Semantic Fragmentation: Static chunking processes document blocks independently, losing or duplicating concepts that span chunk boundaries.The resulting graph lacks the density required for reliable multi-hop reasoning.
- Rigid Parallelism: Graph-RAG ingestion is constrained by sequential processing or fixed parallelism, causing long indexing times or OOM risks on heterogeneous hardware.The introduction reports no existing pipeline with runtime concurrency adaptation, persistent cross-worker state, and crash recovery.
- Cross-Domain Isolation: Multiple specialized knowledge bases require either merging domains, which dilutes specialization and pollutes context, or manual user routing.No existing Graph-RAG system is identified as discovering implicit structural connections between separate graphs at query time.
- Contributions: Noesis addresses these bottlenecks with bidirectional graph traversal, AIMD concurrency control, domain-aware MoE quantization, and Mesh cross-KB routing.The contributions include Graph-Feedback context resolution, Redis-backed persistent state, runtime re-adaptation, and runtime structural discovery with adaptive Natural Break thresholds.
2 System Architecture
Noesis uses five decoupled components connected through HTTP REST APIs and Redis message queues, enabling independent scaling and backend replacement. The architecture separates inference, extraction, orchestration, multimodal ingestion, and cross-knowledge-base routing.
- Component Architecture: Five decoupled components communicate through REST APIs and Redis message queues, enabling independent scaling and backend replacement.The architecture comprises Cortex, Neuron, Synapsis, Retina, and Mesh.
- Component Architecture: Cortex provides independently scalable LLM inference on dedicated GPUs, using llama.cpp with Mo¯esis optimization.
- Component Architecture: Neuron extracts knowledge graphs through bidirectional traversal, entity deduplication, and cross-linking.It operates across textual corpora including documents, web pages, transcriptions, and software repositories without pipeline modifications.
- Component Architecture: Synapsis orchestrates distributed jobs through an AIMD controller and API gateway, while Retina ingests audio and video using C++ VAD and speech-to-text.
- Component Architecture: Mesh performs cross-KB semantic routing as a shared in-process kernel with < 2 ms latency.
3 Bidirectional Graph Traversal
Noesis uses bidirectional graph traversal to preserve semantic continuity across long documents: a forward pass carries selectively recalled context with degrading memory, while a conditional backward pass reconnects early and late concepts. On multilingual and long-document evaluations, this approach substantially expands graph structure and recovers long-range causal relationships.
- Forward traversal: Noesis processes document slices sequentially, using the Graph-Feedback Context Resolver to select relevant previously extracted nodes for each upcoming slice.The resolver combines unigram and bigram keyword overlap with a recency bonus that decays with document distance.
- Adaptive context: The context budget scales linearly with slice count while remaining bounded by minimum and maximum limits.Semantic boundaries such as headings and paragraph breaks keep slices aligned with complete conceptual units.
- Backward traversal: The backward pass reconnects early low-degree nodes to semantically related late nodes, closing long-range dependencies missed during forward extraction.It activates only for sufficiently long documents with enough extracted nodes to benefit from reconnection.
- Evaluation: 253 nodes, 257 edges, 179 cross-links, and 12 communities result from bidirectional traversal, compared with 29 nodes, 36 edges, and 4 communities for the single-pass baseline.The multilingual knowledge base comprised 5 PDF documents.
4 AIMD Adaptive Parallel Processing
Noesis applies backend-agnostic AIMD control to maximize inter-document ingestion parallelism while preserving the intra-document sequentiality required for bidirectional traversal. The controller adapts concurrency under unknown backend capacity and failures, achieving 23× speedup with zero OOM events and crash-resilient state.
- Design constraints: Bidirectional traversal processes slices sequentially within each document while parallelizing across documents to preserve Graph-Feedback context and cross-section connections.This separates required intra-document ordering from scalable inter-document throughput.
- AIMD controller: AIMD increases document concurrency by one after consecutive successes and halves it after consecutive failures, subject to configured bounds.The controller starts conservatively and adapts at runtime rather than assuming backend capacity.
- AIMD controller: The controller remains backend-agnostic across local GPUs, AWS Bedrock or Claude, and Ollama by decoupling inference from orchestration.It accommodates differing latency profiles, rate limits, transient errors, and capacity characteristics.
- Results: 23× speedup was achieved on 3 PDF documents totaling 13.4 MB, reducing processing time from ∼25 min sequential to 1 min 6 s with adaptive parallelism.The reported comparison is specific to this three-document test.
- Results: Zero OOM events were observed across all test configurations, while controller state survived worker crashes and restarts without a warmup period.These results address runtime stability and recovery under adaptive processing.
5 Mo¯esis: Domain-Aware Selective Quantization
Mo¯esis uses domain profiling to identify hot and cold experts, then applies selective quantization, GPU promotion, and runtime re-adaptation for consumer-hardware inference. It enables operation under extreme VRAM constraints while improving processing speed as more layers fit on the GPU.
- Level 2: Selective Quantization and GPU Promotion: 6-bit precision is assigned to hot layers and 2-bit compression to cold layers, shrinking the model from 21 GB to ∼16 GB on disk.The selective policy combines activation concentration with the frequency of the most-used experts.
- Level 2: Selective Quantization and GPU Promotion: Promote-only placement moves layers from CPU to GPU until VRAM is full, reducing repeated CPU↔GPU transfers without manual configuration.Profiling, classification, quantization, and placement execute automatically; layers are never demoted.
- Level 3: Runtime Re-Adaptation: Runtime re-adaptation re-profiles and re-quantizes from the preserved full-precision model when the domain changes, preventing cumulative precision loss.For example, adaptation can move from medical documents to engineering without chaining quantization errors across cycles.
- Consumer-Hardware Results: 160+ minutes of continuous operation produced zero crashes on 6 GB VRAM, where Mo¯esis is required for the system to start.The 6 GB setup is a laboratory stress test demonstrating stability, not a recommended operational configuration.
6 Mesh: Cross-KB Routing and Discovery
Mesh routes queries across separate knowledge bases and discovers cross-domain connections at runtime, avoiding both monolithic context pollution and user-specified domain routing. Its selective discovery enables a 2.3B-parameter model to perform multi-hop cross-domain reasoning while rejecting irrelevant connections.
- Cross-KB Routing and Discovery: Mesh addresses monolithic context pollution and manual domain routing through automatic cross-KB routing with runtime structural discovery.Queries are routed to relevant knowledge bases via hierarchical fingerprint matching, after which emergent connections are identified across separate graphs.
- Cross-KB Routing and Discovery: Queries are routed primarily by cosine similarity to semantic KB fingerprints, with community-overlap and lexical-label fallbacks, then fused through five retrieval phases.Selected knowledge bases are queried in parallel; fusion performs extraction, confidence scoring, Jaccard n-gram deduplication, reranking, and a top-k cap.
- Cross-KB Routing and Discovery: Mesh discovers emergent links through chunk-level semantic similarity and node-level graph comparison using embeddings plus structural bonuses from degree and community membership.Node-level discovery expands query-relevant nodes to one-hop neighbors before comparing nodes across selected knowledge bases.
- Cross-KB Routing and Discovery: Adaptive knee-based thresholds lower for semantically close knowledge bases and remain high for distant domains, with percentile fallback and a calibrated safety clamp.The threshold is computed per query from sorted cross-KB similarities and their first differences rather than using one fixed global value.
- Evaluation Example: A 2.3B-parameter model connected cortisol, endocrine effects, emotional regulation, and couple communication with Mesh, including the emergent “physical contact → oxytocin → couple bonding” link.Without Mesh, the response was fragmented and acknowledged incomplete information; with Mesh discovery and Natural Break, it was confident and included mitigation strategies without disclaimers.
- Evaluation and Implementation: Mesh produced zero cross-KB connections for a software-codebase knowledge base against semantically unrelated domains, while in-process execution achieved routing latency <2 ms.These results demonstrate selective rather than indiscriminate routing and avoid the network round-trips of external service calls.
7 Evaluation
Noesis achieves strong multi-hop QA performance despite using an on-premises graph-construction model, and its retrieval architecture remains effective with a 2.3B answer model and lower retrieval budgets. Evaluation on source code further shows source-agnostic reasoning across modules without pipeline changes.
- Model-Scale Ablation: EM=47.20 and F1=60.30 with a 2.3B answer model match dense retrieval’s EM=47.60 and F1=60.36 with a model approximately 100× larger.The Noesis graph, built by Qwen3.6-35B-A3B, is identical across configurations, isolating retrieval architecture from answer-generation model scale.
- Retrieval-Budget Ablation: EM=60.90 and F1=76.12 at k=20 show that doubling Noesis’s retrieval budget yields only +1.4 EM.At matched k=20, Noesis matches HopRAG in F1, 76.12 vs. 76.06, while retaining on-premises graph construction.
- Source-Agnostic Evaluation: The source-code evaluation produces a 978-node, 3,824-edge graph and correctly traces an 8-step execution pipeline spanning 5+ modules.The pipeline runs from HTTP endpoint through orchestration, parallel KB queries, result fusion, cross-KB discovery, and streaming response without pipeline modifications.
8 Related Work
Prior Graph-RAG, concurrency-control, MoE-quantization, and multi-KB-routing systems address related challenges but leave gaps in cross-document continuity, document-level sequential extraction, domain-adaptive quantization, or runtime connection discovery. Noesis distinguishes itself through bidirectional document-granular control and domain-specific quantization integrated with cross-KB semantic discovery.
- Graph-RAG Systems: GraphRAG, LightRAG, LazyGraphRAG, HippoRAG, and CrossAug leave limitations in static or independent chunking, cross-document continuity, or dense inter-document graph construction.GraphRAG uses static chunking and sequential processing; LightRAG and LazyGraphRAG do not resolve cross-document continuity, while HippoRAG extracts simplified triples without dense inter-document graphs.
- LLM Concurrency Control: AIMD systems such as CONCUR and HiveMind regulate concurrent LLM workloads at GPU KV-cache or HTTP-proxy levels for unconstrained parallel batch serving.These systems allow requests to be processed in parallel with one another, unlike Noesis’s intra-document ordering constraint.
- LLM Concurrency Control: Noesis’s controller operates at document granularity because bidirectional traversal requires ordered intra-document slices to construct Graph-Feedback context for the backward pass.The controller must also adapt to backends with unknown capacity; the passage states this is a fundamentally different problem from token-, KV-cache-, or agent-level control.
- MoE Optimization: Moesis derives quantization from domain-representative expert activations, uses promote-only GPU placement, and retains full-precision weights for runtime re-adaptation across deployment domains.Its integration targets the full quantization lifecycle and avoids repeated CPU↔GPU transfers after selective quantization.
- Multi-KB Routing: R1-Router, DAKS, HydraRAG, and Adaptive-k address multi-KB routing, cross-source evidence fusion, verification, or retrieval quantity selection rather than runtime connection discovery.Adaptive-k uses largest-gap detection for retrieval quantity, which the passage explicitly distinguishes from connection-discovery thresholds.
9 Limitations and Future Work
The paper identifies hardware-, corpus-, and backend-dependent performance measurements, limited evaluation breadth, and incomplete characterization of Mesh across query types and knowledge-base configurations. Future work should broaden benchmarks, domains, queries, and KB setups to better define Noesis’s boundaries.
- Performance limitations: 23× for AIMD parallelism and 6.3× for Mo¯esis prompt processing depend on the tested hardware and corpus, though qualitative trends are expected to hold.Absolute speedups may vary across corpora and backend configurations.
- Evaluation scope: Evaluation spans HotpotQA, two long documents, a software codebase, and three cross-domain knowledge bases, but broader benchmarks and domains remain necessary.Additional evaluation would further characterize the system’s boundaries.
- Mesh evaluation: Mesh routing is demonstrated on a representative query across three domain-specific knowledge bases, leaving wider query types and KB configurations to characterize.The current evaluation establishes the mechanism but does not cover the full range of cross-KB conditions.
10 Conclusion
Noesis is a fully implemented Graph-RAG system introducing four algorithmic innovations, including bidirectional traversal with Graph-Feedback context. It was extensively tested across heterogeneous, multilingual corpora and achieved 90% source-verified precision on long-range causal edges while surpassing GraphRAG by +27.8 EM on HotpotQA.
- System contributions: Noesis introduces four algorithmic innovations addressing persistent limitations in existing Graph-RAG approaches.The system is described as fully implemented.
- Graph construction: 90% source-verified precision was achieved on long-range causal edges extracted from a 193-page document.Bidirectional traversal with Graph-Feedback context produced denser knowledge graphs than independent chunking.
- Evaluation scope: The system was tested on more than 60 documents and 170 MB across multiple knowledge bases, source types, document formats, and languages.The test corpus included PDF, DOCX, a complete software codebase, non-Latin scripts, and documents approaching 200 pages.
- Retrieval evaluation: +27.8 EM over GraphRAG was demonstrated on HotpotQA using a 35B on-premises model.The result is reported as competitive multi-hop retrieval quality in the quantitative evaluation.