← Back to Paper Radar

Blackhole AI: Intelligent Query Routing and Cost-Optimized RAG System

Mr. Sahil A. Katkar, Mr. Harshal N. Barge, Mrs. Kajal P. Khalate, Mr. Prathmesh V. Ingale, M. Yeske
International Journal for Research in Applied Science and Engineering Technology | 2026
Blackhole AI is a query routing framework that computes semantic embeddings and a quantitative complexity score for each incoming query to decide whether it should be handled by a cheap local model or a more capable cloud LLM, using retrieval-augmented generation to help preserve answer quality across both paths.

Problem Statement

Most production LLM systems route every query through the same large cloud-hosted model regardless of how simple or complex it is, which wastes compute, increases latency, and limits scalability for enterprise knowledge applications. This work targets that inefficiency by estimating query difficulty before invocation so that computational resources can be allocated proportionally to actual task complexity, aiming for more economically sustainable LLM deployment.

Key Novelty

  • A cost-aware routing layer that scores query complexity upfront (pre-generation) rather than relying on post-hoc quality checks
  • Joint use of semantic embeddings for both RAG retrieval and complexity/difficulty estimation within a single routing pipeline
  • Adaptive decision thresholds enabling dynamic switching between local and cloud models rather than a static or rule-based routing policy

Evaluation Highlights

  • No concrete quantitative benchmarks (e.g., cost savings %, latency reduction, accuracy retention) are reported in the abstract
  • Claims of balancing accuracy and operational efficiency are presented conceptually/qualitatively rather than validated against baseline systems or public datasets

Signal Assessment

2/10 The system combines already well-established techniques (semantic embeddings, RAG, complexity-based model routing) in a manner conceptually similar to prior LLM cascading/routing work (e.g., FrugalGPT, RouteLLM), without a demonstrably new algorithmic mechanism or reported quantitative benchmark results.

Methodology

  1. Encode each incoming query into a semantic embedding vector
  2. Estimate query complexity/difficulty using a quantitative scoring function
  3. Apply a cost-aware routing decision with adaptive thresholds to select a local or cloud-based model
  4. Retrieve relevant context via vector-based RAG retrieval to augment the selected model's generation
  5. Generate and return the response, balancing accuracy against compute cost

System Components

Semantic Embedding Module

Converts queries into vector representations used for both retrieval and complexity assessment

Complexity Estimator

Quantitatively scores query difficulty to inform which model tier should handle the request

Cost-Aware Router

Uses adaptive decision thresholds to dynamically dispatch queries between local and cloud models

RAG Retrieval Engine

Performs vector-based retrieval of relevant documents/context to ground generated responses

Local Model Backend

Handles low-complexity queries at reduced latency and cost

Cloud Model Backend

Handles high-complexity queries requiring stronger generative capability

Results

Metric Baseline (Uniform Cloud Routing) Blackhole AI Delta
Inference Cost All queries sent to cloud model Simple queries offloaded to local model Qualitative reduction claimed; no % reported
Latency Uniform cloud-bound latency for all queries Lower latency for low-complexity queries Qualitative improvement claimed; no figures reported
Response Accuracy Cloud-model baseline accuracy Claimed to be preserved via adaptive thresholds Not quantitatively verified in abstract

Key Takeaways

  • Complexity-aware routing is a practical lever for cutting LLM serving costs in systems with heterogeneous query difficulty
  • Pairing RAG retrieval signals with routing logic can help avoid defaulting to large models simply to compensate for missing context
  • Before adoption, practitioners should demand quantitative cost/latency/accuracy tradeoff data and comparisons against established routing baselines like FrugalGPT or RouteLLM

Abstract

The rapid adoption of large language models (LLMs) in enterprise and knowledge-intensive applications has introduced significant challenges related to inference cost, latency, and scalability. Most existing deployments rely on uniform cloud-based processing, which leads to unnecessary resource consumption for simple queries. This paper presents Blackhole AI, an adaptive query routing framework that integrates seman-tic embedding, retrieval-augmented generation, and quantitative complexity estimation to dynamically select between local and cloud-based models. The proposed system introduces a cost-aware routing mech-anism that evaluates query difficulty before model invocation, enabling efficient allocation of computational resources. By com-bining vector-based retrieval with adaptive decision thresholds, Blackhole AI aims to balance response accuracy with operational efficiency. The framework highlights the importance of intelligent routing strategies in achieving scalable and economically sustain-able LLM deployment.

Generated from available metadata and abstract on 2026-07-14 using Claude.