RAG Signal™ Adaptive RAG Architecture

1. Introduction

The rapid proliferation of generative AI systems has fundamentally reshaped the digital information landscape. Users increasingly rely on synthesized answers from these models rather than ordered lists of links. This paradigm shift makes brand representation in LLM-generated content a critical competitive concern. Traditional Search Engine Optimization (SEO) is inadequate for this new reality, leading to the emergence of Generative Engine Optimization (GEO).

Most production RAG deployments exhibit four critical limitations that the RAG Signal architecture directly addresses:

1. Static indices: No automatic updates when content changes — resolved via djb2 content-hash-based incremental indexing.
2. Single-source dependency: Only site content is indexed — resolved via multi-source ingestion with differential weighting.
3. Task-agnostic retrieval: The sole ranking criterion is cosine similarity; task context is ignored — resolved via task-oriented query enrichment.
4. Temporal blindness: Old and new data are weighted equally — resolved via rational decay freshness scoring with a 180-day hard cutoff.

2. Related Work

Retrieval Augmented Generation (RAG) systems have become a significant research area for addressing the issues of information currency and accuracy in Large Language Models (LLMs). While traditional RAG approaches often rely on static knowledge bases and simple similarity metrics, recent work has focused on overcoming these limitations.

Adaptive RAG approaches, such as SPARC-RAG, aim to dynamically adjust retrieval strategies based on query context. Temporal RAG, exemplified by STAR-RAG and Chronofy, integrates recency-aware retrieval into RAG pipelines. RAG Signal extends these paradigms with multi-source differential weighting, task-aware enrichment, and a production-hardened temporal freshness function with a hard cutoff threshold.

3. System Architecture

3.1 Robust Hash-Based Incremental Updates

A djb2 content hash of the raw page content serves as a fingerprint, enabling incremental re-indexing. This triggers re-indexing only when structural or partial content changes occur, saving embedding API costs and database writes for unchanged content.

3.2 Vector Representation

The primary embedding model produces 1,024-dimensional vectors via Jina AI's jina-embeddings-v3, stored in PostgreSQL with pgvector HNSW indexing for sub-millisecond approximate nearest neighbor search.

3.3 Normalized Similarity with ReLU Clamping

Semantic similarity is clamped using a ReLU filter bounded between 0 and 1, preventing negative cosine similarities from producing meaningless negative scores when multiplied with source weights and freshness factors: s'(q, c) = max(0, sim(q, c)).

3.4 Differential Source Weighting

3.5 Temporal Freshness Scoring with Hard Cutoff

Data recency is quantified via a rational decay function: f(Δt) = 1 / (1 + Δt/30), where Δt is days since last update. At 180 days, the score is zeroed out — preventing stale data from persisting indefinitely at diminishing non-zero weights.

3.6 Task-Oriented Query Enrichment

Raw user queries are augmented with task-specific keyword suffixes before embedding, improving retrieval relevance: meta → "page title description content keywords", discovery → "services audience USP keywords brand positioning", audit → "technical issues recommendations content quality", hallucination → "facts services team location founding year", score → "content structure readability entities schema".

3.7 Composite Ranking Score

The mathematical pre-ranking score is the product of three independent factors:

score(c) = max(0, sim(q, c)) × w(source(c)) × f(updated_at(c))

The top 2k candidates (where k is the desired result count) proceed to LLM re-ranking.

3.8 Hybrid Ranking: Semantic + Keyword Fusion

Retrieval uses a 70/30 semantic-to-keyword weighted hybrid approach via PostgreSQL, ensuring both conceptually relevant and term-exact content is surfaced.

3.9 LLM Re-Ranking

After mathematical pre-ranking, an LLM-based re-ranker (DeepSeek Reasoner) performs contextual relevance assessment. On failure, the system gracefully degrades to mathematical ranking, ensuring production reliability.

3.10 KB Curator Agent

An autonomous curation pipeline classifies each ingested chunk as keep, improve, or delete. Navigation menus, cookie notices, and boilerplate are removed. Improved content is re-embedded with fresh vectors. The curator is idempotent — only uncurated content is processed.

4. Implementation

The RAG Signal™ Adaptive RAG architecture is implemented as a set of Supabase Edge Functions (TypeScript/Deno), open source under Apache 2.0:

5. Experimental Results

The architecture has been validated across ten production client deployments representing seven industry sectors:

63–105 prompts per deployment. Attribution measured via Mistral Large with web search, evaluating brand presence in response text and citations.

6. Conclusion

The RAG Signal™ Adaptive RAG architecture successfully addresses the four core limitations of traditional RAG systems through djb2 hash-based incremental updates, multi-source differential weighting, ReLU-clamped normalized similarity, task-oriented query enrichment, temporal freshness with 180-day hard cutoff, hybrid semantic-keyword fusion, and autonomous KB curation. With a 77.1% average attribution rate across ten production deployments, the architecture provides a robust, validated framework for the next generation of Generative Engine Optimization.

References

1. djb2 Hash Function. (Accessed: July 26, 2026).
2. Rectified Linear Unit (ReLU) Activation Function. (Accessed: July 26, 2026).
3. Exponential Decay Function. (Accessed: July 26, 2026).
4. O. Martinez, "Optimizing Visibility in Generative Engines," arXiv:2607.14035, 2026.
5. Z. Tian et al., "Diagnosing and Repairing Citation Failures in GEO," arXiv:2603.09296, 2025.
6. "What is RAG? Latest Advances," Squirro Blog, 2026.
7. "How RAG Works for Enterprise AI," Techment Blog, 2026.
8. "SPARC-RAG," arXiv:2602.00083, 2026.
9. Z. Zhu et al., "STAR-RAG," arXiv:2510.16715, 2025.
10. "Agentic RAG: A Survey," arXiv:2501.09136v4, 2026.
11. "GEO Guide 2026," Digital Applied Blog, 2026.
12. "GEO: The 2026 Guide," LLMRefs, 2026.
13. "GEO Statistics (2026)," Omnibound AI Blog, 2026.
14. "Mastering GEO in 2026," Search Engine Land, 2026.
15. M. Syed et al., "Chronofy," arXiv:2607.20560, 2026.