Jev with RAG introduces a new layer of decision-making to the classic Retrieval-Augmented Generation pipeline, allowing systems to filter candidates not just by similarity but by their actual ability to answer the query.
Jev with RAG Overview
The Jev + RAG combination preserves the initial retrieval step of a standard RAG system but inserts a novel gating phase before generation. In this phase the original query is treated as a shared state, while each retrieved passage becomes an independent candidate. Jev evaluates every candidate against a typed question such as "Đoạn văn này có giúp trả lời câu truy vấn không?" and produces a probability that reflects how well the passage can answer the query.
Standard RAG Workflow
A conventional RAG pipeline first splits source documents into manageable chunks. Each chunk is transformed into a dense vector via an embedding model and stored in a vector database. When a user query arrives, the same embedding model converts the query into a vector, and the system retrieves the top-k nearest chunks. Many production deployments add a reranker that compares the query with each retrieved passage, reorders them, and keeps the highest-scoring items. The selected passages are then placed into the LLM's context window for answer generation.
Ranking vs. Answerability
It is important to distinguish between ranking-how closely a passage matches the query vector-and answerability-whether the passage actually contains sufficient evidence to produce a correct answer. A passage may rank highly yet provide only weak or incomplete evidence, leading the LLM to generate a plausible-looking but unsupported response.
Replacing the Traditional Reranker with Typed Questions
Instead of the usual similarity-based reranker, Jev can assess each retrieved candidate using a clear-typed question. For example, the system asks: "Đoạn văn này có giúp trả lời câu truy vấn không?" This transforms the ranking problem into a binary relevance assessment, yielding a probability score for every candidate. The application then applies a threshold: candidates above the threshold proceed to the LLM, while those below are discarded from the context window.
Threshold Gating and Answerability Check
Within the same request, Jev also evaluates whether the remaining evidence collectively meets a predefined answerability threshold. If the combined answerability score falls short, the application may bypass the LLM entirely and return the message "không có trong tài liệu" ("not found in the documents"). This gating prevents the model from fabricating answers when the source material is insufficient.
Practical Considerations and Limitations
Jev does not replace the embedding model, vector store, or the generation model; retrieval still defines the upper bound of what can be answered because Jev cannot resurrect a passage that was never retrieved. Organizations should monitor the chosen thresholds to balance recall and precision, and be aware that overly aggressive gating may reduce coverage, while lax thresholds may let low-answerability passages influence the LLM.
Conclusion
By integrating Jev with RAG, developers gain fine-grained control over both ranking and answerability, enabling systems to filter out weak evidence before generation and even return a clear "không có trong tài liệu" response when the data does not support an answer. This approach strengthens reliability while preserving the core strengths of the traditional RAG architecture.
