Bottom Line
Federated Search fans every natural-language question out across three complementary retrieval methods — semantic, keyword, and knowledge-graph — and fuses the results into one answer with unified citations. The result is materially better answer quality and breadth than any single retrieval method delivers alone.
How It Works: Three Modalities, One Answer
A single question is fanned out across three retrieval engines simultaneously, each contributing what it's best at:
| Modality | Store | Strength It Contributes |
|---|---|---|
| Semantic | Dense vectors (Qdrant) | Meaning and paraphrase — finds relevant content even when the wording doesn't match, fuzzy recall |
| Lexical | Elasticsearch (BM25) | Exact terms, identifiers, and rare tokens — precision on the specific phrase or code a user actually typed |
| Graph | Neo4j (knowledge graph) | Entity relationships, multi-hop reasoning, and provenance traversal — connects facts across documents, not just within one |
The three result sets are combined using reciprocal-rank-fusion and a cross-modality reranker, and the final answer carries unified citations spanning all three modalities — so a user gets one coherent, sourced answer regardless of which retrieval method actually found the supporting evidence.
Why It Matters
- No single retrieval method is good at everything. Semantic search alone misses exact identifiers and codes; keyword search alone misses paraphrased or conceptually related content; graph search alone misses free-text nuance. Fusing all three closes the gaps that show up as "the AI missed something obvious" or "the AI didn't connect the dots."
- This is the answer to buyers who ask why the platform doesn't just rely on vector search — the platform's position is that meaning, keyword, and relationship retrieval are complementary, not competing, and combining them is measurably better than any one of them alone.
- The graph modality in particular unlocks answers that require connecting facts across multiple documents — entity relationships and multi-hop reasoning — which a single-document retrieval pass cannot surface on its own.
Trust Is Preserved Across All Three Channels
Adding two new retrieval stores (Elasticsearch and Neo4j) does not create two new places where governance can be skipped. Each new store is brought into the same encryption, access-control, and audit scope as it lands, and the platform's core access rule — a user can only retrieve content at or below their clearance level — is enforced identically across all three modalities. Unified citations preserve the same source, trust, and classification provenance regardless of which channel produced the result, so there is no path for a keyword or graph-relationship result to leak content a user shouldn't see.