WUF.AI
Framework ID: RAG-X
Ver 2.1.0

Precision
Retrieval.

Go beyond basic vector search. WUF.AI utilizes a multi-stage retrieval pipeline with HyDE, hybrid search, and cross-encoder reranking to guarantee the highest relevance for every query.

Avg Latency
142msOPTIMAL
Recall@10
0.984STABLE
NDCG@5
0.921OPTIMAL
Cache Hit Rate
42.5%WARNING
Start Audit

HyDE Generation

We use Hypothetical Document Embeddings (HyDE) to generate a theoretical answer first, mapping the query closer to the actual document space before searching.

Hybrid Search

Combining dense vector search (semantic meaning) with sparse keyword search (BM25) to capture both concepts and exact terminology.

Cross-Encoder

Top results are passed through a powerful cross-encoder model to score the exact relationship between the query and the retrieved chunks.

Retrieval Architecture

Multi-Stage Pipeline

Query
HyDE
Vector Search
Cross-Encoder
PIPELINE_STATUS
QUERY STAGE ACTIVE
ARCHITECTURE

Hierarchical Chunking

Standard RAG breaks documents into arbitrary sizes, destroying context. WUF.AI uses semantic boundaries and a parent-child relationship model.

  • Semantic BoundariesWe split text at logical breaks (headers, paragraphs, code blocks) rather than arbitrary character counts.
  • Parent-Child RetrievalWe embed small, precise "child" chunks for accurate search, but retrieve the larger "parent" chunk to provide the LLM with full context.
Parent-Child Chunking
halfvec(768)
PARENT (RETRIEVED)
"The Enterprise plan includes SSO, audit logs, and dedicated support. Pricing starts at $499/month for up to 50 users. Additional users are billed at $10/user/month."
CHILD 1 (EMBEDDED)
"The Enterprise plan includes SSO, audit logs, and dedicated support."
CHILD 2 (MATCH)
"Pricing starts at $499/month for up to 50 users."
Query: "How much is enterprise?"
Similarity: 0.92
query_builder
SELECT chunk_id, text, similarity
FROM document_chunks
WHERE similarity > 0.85
AND document.source_type = 'github'
AND document.metadata->>'status' = 'merged'
AND document.created_at > '2023-01-01'
ORDER BY similarity DESC
LIMIT 10;
PRECISION

Pre-Filtering

Vector search alone is noisy. WUF.AI applies strict metadata filters before the vector search executes, drastically reducing the search space and eliminating irrelevant results.

  • Attribute QueriesFilter by source type (e.g., only Slack), author, date range, or custom tags applied during ingestion.
  • ABAC EnforcementPre-filtering is also how we enforce security. If a user doesn't have access to a document, its vectors are excluded from the search space entirely.
PERFORMANCE

Semantic Caching

Why compute the same answer twice? WUF.AI caches responses based on semantic similarity, not just exact string matches.

  • Sub-100ms ResponsesIf a user asks "What is the enterprise price?" and another asks "How much for enterprise?", the cache serves the identical answer instantly.
  • Cost ReductionBy serving answers from the cache, you bypass the LLM generation step entirely, saving significant API costs on high-volume queries.
Cache Hit42ms
QUERY 1 (CACHE MISS)
"What is the SLA for the Enterprise plan?"
LLM Generation: 2.4s
QUERY 2 (CACHE HIT)
"Tell me the enterprise SLA."
Semantic Match (0.94) - Served from Cache
ACCURACY

Cross-Encoder Reranking

Bi-encoders are fast but imprecise. Cross-encoders are slow but highly accurate. We use both.

1. Broad Retrieval

We use fast bi-encoder vector search to retrieve the top 100 potentially relevant chunks from millions of documents in milliseconds.

2. Deep Scoring

A cross-encoder model evaluates the query and each chunk simultaneously, understanding the deep semantic relationship between them.

3. Final Selection

Only the top 5-10 highest-scoring chunks are passed to the generative LLM, ensuring it only sees the most relevant, high-signal context.

RAG & Search FAQ

What is HyDE and why do you use it?

HyDE (Hypothetical Document Embeddings) improves search accuracy by generating a theoretical answer to a query first. We embed this hypothetical answer and use it to search the vector database, which often yields better matches than embedding the raw, short user query.

How does parent-child chunking improve context?

Standard chunking often cuts off important surrounding context. We embed small "child" chunks for precise semantic matching, but when a match is found, we retrieve the larger "parent" chunk (e.g., the whole paragraph or section) to give the LLM the full context it needs to generate a complete answer.

Does semantic caching compromise security?

No. Semantic cache hits are strictly scoped by the user's ABAC permissions. A user will never receive a cached answer generated from documents they do not have explicit permission to access.

Stop Hallucinating.

Experience the difference of a precision-engineered retrieval pipeline.

Book a Custom Demo