If a model generates responses that include information not found in the retrieved context, it introduces closed-domain hallucinations. This means the model is making up facts rather than relying on retrieved information, leading to misinformation and reduced trust.

​
Example of the Problem

User Query: “What year was the Eiffel Tower completed?”

Retrieved Context: “The Eiffel Tower is an iron lattice tower located in Paris, France. It was designed by Gustave Eiffel.”

Model Response: “The Eiffel Tower was completed in 1889 and is the most visited paid monument in the world.”

​
What Went Wrong?

  • What We Did Wrong:
    • Retrieved documents contained irrelevant information.
    • The model overgeneralized or extrapolated beyond what was retrieved.
    • The retrieval pipeline was returning too many noisy or loosely related chunks.
  • How It Showed Up in Metrics:
    • Low Context Adherence: The model included information not present in the retrieved documents.
    • High Chunk Attribution but Low Chunk Utilization: The model referenced retrieved data but incorporated only small portions of it.

​
Improvements and Solutions

Skim through each of these solutions before choosing the best one for your situation!

1

Enforce Context Adherence in Prompts

  • Being more explicit in system prompts can help it stick to just the facts provided: 
    Instruction: Only use the retrieved context to answer the question. If the answer is not found, state 'I don’t know.'
  • You can also modify the prompt to recommend it uses more of the provided context if there are many relevant results. 
    Instruction: Use at least 3 sources in your answer if they are relevant.
2

User Queries Are Too Vague

Use query expansion techniques to reformulate the user’s query to retrieve more relevant context.

Ex. User Query: “Eifel Tower” -> “What is the Eifel Tower?”, “When was the Eifel Tower created?”, etc.

  • You can use an LLM on their query to “guess” the right answer, then use that guess as part of your context retrieval search.
  • You can also generate multiple “alternate” queries to see if they find more relevant context.
  • For short queries, you can expand them into relevant questions.
3

Chunks Are Truncated

  • Increase the chunk size to prevent truncation of important details.
  • Apply better chunking strategies to ensure context retrieval is more structured and relevant.
    • Ex. Use overlapping sliding windows to maintain continuity in extracted information.
4

Retrieved Chunks Are Irrelevant

  • Switch to a more powerful embedding model for retrieval to improve similarity matching.
  • Implementing re-ranking algorithms to prioritize the most relevant chunks from retrieved data. This can give better adherence, but with higher compute required.
    • To mitigate the performance cost of cross-encoding, you can use the faster bi-encoding method to get an initial pool to prioritize further.
  • Increase similarity thresholds to eliminate loosely related retrieved data.
5

Use Hybrid Retrieval Techniques

  • Combine dense vector search (embeddings) with sparse retrieval (BM25, TF-IDF) for better precision.
    • This can improve Context Adherence and Chunk Utilization.
6

Fine Tune For Adherence

  • Penalize generations that introduce out-of-context information by fine-tuning the model with contrastive learning techniques.
  • For some RAG setups, you may accomplish higher adherence by tuning meta-parameters when generating the response (ex. reducing temperature).
    • This can improve Context Adherence.
7

Validate Responses for Adherence

  • Use Context Adherence Plus to generate explanations for why responses are not contextually aligned.
  • Flag responses with adherence scores below a threshold for human review.
  • Apply post-processing filters to remove non-contextual information before presenting responses to users.