SaSame Research Agent
RAG reduces hallucination but weak retrieval, noisy context, and unconstrained generation still cause errors. Key fixes span hybrid search, reranking, citation prompting, and faithfulness evaluation.
Retrieval-augmented generation inserts relevant documents into the model's context before generation, giving it factual material to draw from. Yet retrieval is imperfect: top-k passages may be partially relevant, stale, or too numerous to process cleanly. When the retrieved context does not fully answer the query, the generator fills gaps with parametric memory — knowledge baked into its weights — and this is where hallucination enters. Improving retrieval is therefore the highest-leverage starting point, upstream of any prompt-level fix.
Hybrid search addresses retrieval recall by combining vector and sparse retrieval, then merging results via reciprocal rank fusion or a learned combiner. This pairing captures paraphrase-heavy matches and exact-term matches that a single method would miss. Reranking then applies a cross-encoder to the merged candidate set, reordering by fine-grained relevance and discarding low-signal passages before they reach the generator. Together, hybrid retrieval and reranking narrow the gap between what is retrieved and what is actually useful for grounded generation.
On the generation side, prompt-level constraints are the simplest intervention with broad impact. Instructing the model to cite a specific passage for each claim — and to respond with a structured 'not enough information' signal when no passage supports the answer — prevents the model from inventing plausible-sounding details. Agent-native systems, including AI studios like SaSame that build on MCP tool calls and retrieval pipelines, can layer this further by routing citation-unsupported claims to a fallback retrieval step rather than passing them through to the user.
Automated evaluation closes the loop between development and production. Faithfulness-oriented evaluation checks whether generated claims are entailed by retrieved passages, providing a scalar signal for regression testing and live monitoring. Running this evaluation on a representative sample during development identifies which query types or document types drive hallucination most, making targeted fixes tractable rather than speculative. For production systems, a faithfulness score below a defined threshold converts a silent failure — a wrong answer delivered with confidence — into an observable, actionable event that can trigger a retry, a broader retrieval pass, or a human review.
Why do RAG systems still hallucinate despite retrieving relevant documents?
Retrieved passages may be partially relevant, outdated, or too numerous for the model to process cleanly. When the context does not fully answer a query, the generator fills gaps from parametric memory — facts baked into its weights — which is where confabulation enters. Poor chunking, where passages are too large or too small, amplifies this problem by diluting signal with noise.
What is citation grounding and how does it reduce hallucination?
Citation grounding instructs the model to attribute each claim to a specific retrieved passage and to decline if no passage supports the claim. This constraint keeps the model within retrieved evidence rather than relying on memorized priors. It also makes errors auditable: a downstream system or user can verify the cited source directly.
How does hybrid search reduce retrieval-side hallucination risk?
Hybrid search combines dense vector retrieval (semantic similarity) with sparse keyword retrieval (BM25-style). Dense retrieval captures meaning but can miss exact terms; sparse retrieval catches precise keywords but misses paraphrase. Combining both lowers the chance of failing to retrieve a relevant passage that the generator would otherwise fabricate.
What is reranking and when should it be applied?
Reranking is a second-pass scoring step, typically using a cross-encoder model, applied to the initial retrieved candidate set before generation. It reorders candidates by fine-grained relevance to the query and discards low-signal passages, keeping the context window tight. This reduces noise without requiring changes to the retrieval corpus or embedding model.
How can faithfulness be measured after generation?
Frameworks such as RAGAS assess faithfulness by checking whether each generated claim is entailed by the retrieved context, using an LLM or NLI model as judge. A low faithfulness score flags answers that go beyond retrieved evidence. In production, a score below a threshold can trigger a retry with a tighter prompt, a broader retrieval query, or a human-review flag.