Retrieval Technology
Induced-Fit Retrieval
Retrieval that thinks while it searches.
+0.0%
R@5 vs best RAG
0.0M
articles tested
0ms
beam traversal
$0
per query
p = 0.0002 · Zero training · Any embedding model
The Problem
RAG treats every query as a single-shot lookup
Retrieve-Augmented Generation embeds a query, finds the nearest vectors, and returns results. This works when the answer lives in one place.
It fails when the answer requires connecting information across multiple documents — the kind of reasoning humans do effortlessly. Multi-hop questions are not edge cases.
In enterprise knowledge bases, legal discovery, medical research, and technical estimation, the most valuable answers connect disparate facts. RAG cannot follow these chains because it commits to a fixed query before it knows what it needs to find.
How It Works
The Induced Fit Principle
In 1958, Daniel Koshland proposed that enzymes change shape upon contact with their substrate. IFR applies the same principle to information retrieval.
Lock and Key
Traditional RAG
The query is a fixed key. The system finds the best static match. If the answer is semantically distant, retrieval fails.
Induced Fit
IFR by Celestix
The query adapts as it encounters information. After visiting each node, the query shifts toward discovered concepts while retaining its original intent through an anchoring mechanism.
Architecture
Three Stages of Retrieval
Entry
The system receives a query, generates its embedding, and identifies an entry point using HNSW indexing. Microseconds regardless of graph size.
Traversal with Induced Fit
CORE INNOVATIONA traversal agent walks the knowledge graph via beam search. At each node, the query representation mutates — adapting to discovered context while anchoring to original intent. An energy budget governs exploration: the agent gains energy on relevant nodes and loses it otherwise, naturally terminating when returns diminish.
Fusion and Ranking
Graph traversal candidates merge with standard nearest-neighbor results via reciprocal rank fusion. An optional cross-encoder re-ranking stage produces the final result set.
Benchmarks
Validated on 5.2 Million Wikipedia Articles
HotpotQA fullwiki · 500 queries · Bootstrap B=10,000
| Method | R@5 | R@10 | R@20 | MRR | Latency |
|---|---|---|---|---|---|
| RAG top-5 (cosine) | 29.0% | 29.0% | 29.0% | 0.443 | 0.3ms |
| RAG-k20 (cosine) | 29.0% | 32.0% | 34.5% | — | 0.3ms |
| RAG + Cross-Encoder | 33.7% | 33.7% | 33.7% | 0.548 | 16.9ms |
| IFR Beam 100hCelestix | 30.9% | 35.5% | 37.6% | 0.475 | 9.7ms |
| IFR Hybrid+CE 100hCelestix | 36.6% | 36.6% | 36.6% | 0.554 | 24.1ms |
+0.0%
Overall R@5
36.6% vs 33.7%, p=0.0002
+0.0%
Multi-hop R@5
51.0% vs 46.5%, p<0.001
+0.0%
Traversal Reach R@10
35.5% vs 29.0%, p<0.0001
HotpotQA fullwiki, full English Wikipedia (5,233,329 articles). 500 questions (400 bridge + 100 comparison). 21.2M edges. Bootstrap B=10,000.
Scaling
IFR's Advantage Grows with Data
| Corpus | Documents | IFR-beam vs RAG-k5 | Hybrid+CE vs RAG-rerank | p-value |
|---|---|---|---|---|
| FCIS | 722 | −3.7% | +1.4% | n/s |
| MuSiQue | 21,000 | −6.3% | −0.4% | n/s |
| HotpotQA | 66,000 | +1.1% | +3.0% | <0.001 |
| HotpotQA | 508,000 | +2.2% | +4.5% | <0.001 |
| HotpotQA (full) | 5,233,329 | +1.9% | +2.9% | =0.0002 |
Hybrid+CE Advantage by Corpus Size
IFR has no advantage at <20K documents. At 66K+ it becomes statistically significant. The larger your knowledge base, the more IFR matters.
Resilience
17.5M Noisy Edges. Zero Degradation.
Without Hyperlinks
3.7M edges
With Hyperlinks
Improved21.2M edges (+17.5M noisy links)
Pour in millions of messy documents with dirty links. IFR doesn't break. The induced-fit mechanism naturally filters noise during traversal.
Landscape
How IFR Compares
| Category | Players | Method | Multi-hop | Cost |
|---|---|---|---|---|
| Tech Giants | Google, Meta, Microsoft, Apple | RAG + cross-encoder + long context | No (expand window to 2M) | $$$$ GPU clusters |
| Vector DB | Pinecone, Weaviate, Qdrant, Milvus | HNSW / ScaNN / DiskANN | No (single-shot k-NN) | $50–500/mo |
| RAG Frameworks | LangChain, LlamaIndex, Haystack | RAG orchestration + rerank | Iterative (LLM per step) | $$ API per query |
| Agent Memory | Zep, Mem0, Letta | Temporal knowledge graph | BFS (1–2 hops, no mutation) | $$–$$$ SaaS |
| Graph RAG | Microsoft GraphRAG, LightRAG | KG + community detection + BFS | Partial (no query mutation) | $$ LLM for graph build |
| LLM Rewriting | Search-o1, ICR, R1-Searcher | LLM rewrites query → search | Partial (500ms–2s/query) | $$$ LLM call/query |
| IFR | Celestix AI | Graph traversal + induced fit | Yes · 100 hops · 0 LLM calls | $0 locally |
IFR is the only system with mathematical multi-hop query adaptation, zero LLM calls, at sub-10ms latency.
Properties
What Makes IFR Different
Sub-linear Scaling
O(log N) query latency. 52,000× data growth = 3.9× latency growth. Effectively constant for practical corpus sizes.
Zero-shot
No fine-tuning, no training data, no domain adaptation. Point IFR at any pre-embedded corpus and query immediately.
Model-agnostic
Works with any embedding model. Operates on geometric structure, not specific model internals. Swap providers without rebuilding.
Complementary
IFR augments RAG, not replaces it. Fusion architecture combines graph-traversal discoveries with standard k-NN, outperforming either alone.
Deterministic
Every retrieval produces a traversal path — a sequence of nodes and edges explaining how each result was found. Full auditability.
Noise-resilient
+17.5M noisy edges caused zero degradation. Induced fit + beam search naturally filter irrelevant connections during traversal.
Specifications
Technical Details
| Parameter | Value |
|---|---|
| Embedding dimensions | 128D (PCA-compressed from 384D) |
| Graph type | Hierarchical, multi-edge (semantic + cross-reference) |
| Traversal method | Energy-bounded beam search with induced fit |
| Query latency (5.2M nodes) | ~10ms beam, ~31ms with cross-encoder |
| Index build time | Linear in corpus size |
| RAM (Level 0, all nodes) | 48 bytes per atom (~250 MB for 5M atoms) |
| Supported corpus sizes | Tested to 5.2M; designed for 1B+ |
| Cross-encoder | Optional; any BERT-class re-ranker |
| Embedding model | Any (model-agnostic) |
Capacity Model
| RAM Budget | Searchable Atoms | Warm Cache | Active (Full Content) |
|---|---|---|---|
| 1 GB | 3.9M | 500K | 10–50K |
| 8 GB | 31M | 4M | 80–400K |
| 64 GB | 250M | 32M | 640K–3.2M |
Use Cases
Where IFR Applies
Enterprise Knowledge Management
Policies → Procedures → Specifications → Standards
Corporate knowledge is inherently multi-hop.
Legal Discovery
Statutes → Interpretations → Rulings → Amendments
IFR follows citation chains the way researchers think.
Medical Research
Drugs → Pathways → Interactions → Outcomes
Connecting papers, databases, and clinical records.
Construction Estimating
Labor rates → Materials → Geographic factors → Regulations
Where IFR was born. Connecting RSMeans, Davis-Bacon, and contracts.
Customer Support
Documentation → Known issues → Config guides → Past resolutions
Surface the complete picture, not just the closest FAQ.
Intellectual Property
Provisional patent application filed covering:
· Adaptive query mutation during graph traversal
· Energy-bounded beam search with relevance and novelty criteria
· Hierarchical 4-level compressed knowledge graph storage
· Multi-phase edge decay for temporal knowledge management
· Trail persistence for learning from traversal history
Built by Celestix AI
Retrieval that doesn't just search — it reasons.
celestix.ai · Patent pending