Intelligent Retrieval
Retrieval That Thinks Before It Searches
Five layers of intelligence between your question and your answer. Unlike single-strategy systems, Courdx runs semantic search, keyword matching, and knowledge graph traversal simultaneously — then fuses the results into one optimized ranking.
THE FIVE LAYERS
Query Understanding
We Don't Search What You Typed. We Search What You Meant.
Query Decomposition
Example:
"Compare our Q3 revenue to competitors"
What we search:
- 1Our company Q3 revenue figures
- 2Competitor Q3 revenue reports
- 3Revenue gap analysis factors
Predictive Search (HyDE)
Instead of searching for your question, we generate what the perfect answer would look like — then find documents that match THAT. This dramatically improves recall for complex questions.
Query Expansion
Example:
"AR status"
What we search:
- 1Accounts Receivable status
- 2Customer invoice payments
- 3Outstanding balances
Multi-Strategy Retrieval
Three Search Strategies Running Simultaneously. One Intelligent Result.
Semantic Search (Vector)
Understands the meaning behind your words, not just the words themselves. Finds relevant content even when different terminology is used.
Keyword Matching
Exact term matching for precision when it matters — product codes, legal terms, proper nouns, and specific identifiers. Uses BM25, the same algorithm behind traditional search engines.
Knowledge Graph Traversal
Walks the connections between entities to surface related information that neither semantic nor keyword search would find on their own.
Smart Result Fusion (RRF)
Combines results from all three search strategies into one optimized ranking. Each strategy votes on which documents matter most, and the fusion algorithm merges those votes — so you get the best of semantic, keyword, and graph search in every result.
Intelligent Reranking
Finding Documents Is Easy. Finding RELEVANT Documents Is Hard.
Deep Reranking
A specialized AI model reads each query-document pair together and scores true relevance — not just word similarity, but whether the document actually answers your question.
LLM Reranking
For complex queries, an LLM evaluates each result: "Does this actually answer the question?"
Corrective RAG
When Retrieval Fails, We Don't. We Fix It.
Automatic Validation
Every retrieved document graded for relevance. Confidence score calculated before response.
Self-Healing Retrieval
- High confidence → Proceed with answer
- Ambiguous → Refine query, re-retrieve
- Low confidence → Fall back to web search
Guardrails
Protection Built In, Not Bolted On
Input Guardrails
- Prompt injection detection
- PII scrubbing from queries
- Blocked topic enforcement
Output Guardrails
- Hallucination detection
- PII leak prevention
- Citation verification
- Toxicity filtering
GET STARTED
Experience Intelligent Retrieval
See how these five layers work together with your actual documents.