RAG & LLM Generative AI Failures

Definition

Text is split into tokens incorrectly, damaging meaning.

Solution

Use the right tokenizer and test domain-specific terms, names, code, and multilingual text.

Definition

The model struggles with rare words, names, slang, or new terms.

Solution

Use subword tokenization, domain data, updated embeddings, or retrieval.

Definition

The model misses or mislabels names, dates, places, organizations, or products.

Solution

Add better labeled examples and evaluate entity-level precision and recall.

Definition

The model identifies an entity but links it to the wrong real-world concept.

Solution

Use better knowledge bases, disambiguation logic, and context-aware linking.

Definition

The model misunderstands references such as ``he,'' ``she,'' ``they,'' or ``it.''

Solution

Use better context modeling and coreference-specific training or evaluation data.

Definition

The model misses words like ``not,'' ``never,'' or ``without.''

Solution

Add negation-heavy examples and rule-based checks for high-risk tasks.

Definition

The model interprets sarcastic text literally.

Solution

Use domain-specific examples and human review for sensitive decisions.

Definition

The model chooses the wrong meaning for a word or phrase.

Solution

Use more context or ask a clarifying question when the meaning is unclear.

Definition

The model ignores important surrounding text.

Solution

Use better prompts, retrieval, context windows, and targeted evaluation.

Definition

The model assigns the wrong sentiment.

Solution

Use balanced labeled data, domain-specific evaluation, and error analysis.

Definition

The model misunderstands what the user wants.

Solution

Clarify intent labels and train on real user examples.

Definition

The system extracts the wrong values from user input.

Solution

Improve annotation rules, examples, schema validation, and extraction checks.

Definition

The model mistranslates text or loses meaning.

Solution

Use domain-tuned translation and human review for high-risk content.

Definition

A summary includes details not found in the source text.

Solution

Use grounding checks, citation requirements, and source-faithfulness evaluation.

Definition

Generated text does not accurately reflect the input.

Solution

Use extractive checks, citations, factual consistency metrics, and verification prompts.

Definition

Harmful content is missed or safe content is incorrectly flagged.

Solution

Use balanced safety datasets, threshold tuning, and human review for edge cases.

Definition

The model performs worse on dialects, multilingual text, slang, or new terms.

Solution

Continuously evaluate language coverage and update domain data.

Definition

Small wording changes cause large output differences.

Solution

Test prompts across many examples and use robust templates.

Definition

The text is too long for the model to process fully.

Solution

Use chunking, summarization, retrieval, or hierarchical processing.

Definition

The model has access to long text but fails to use the right part.

Solution

Use retrieval, section ranking, summaries, and long-context evaluation.

Definition

The system focuses on matching words instead of meaning.

Solution

Use semantic search, embeddings, hybrid retrieval, and relevance evaluation.

Definition

Text is categorized incorrectly because the model misses the real meaning.

Solution

Improve labels, add representative examples, and evaluate semantic edge cases.

Definition

The model selects too much or too little text as an entity.

Solution

Improve annotation guidelines and use entity-span evaluation.

Definition

Generated text slowly moves away from the original subject.

Solution

Use stronger prompts, section constraints, and validation checks.

Definition

The system fails to retrieve useful documents.

Solution

Improve indexing, embeddings, query rewriting, search strategy, and retrieval evaluation.

Definition

The retriever misses important relevant documents.

Solution

Increase top-k, improve chunking, use hybrid search, and expand queries carefully.

Definition

The retriever returns too many irrelevant documents.

Solution

Use reranking, metadata filters, better embeddings, and relevance thresholds.

Definition

Documents are split in a way that loses meaning.

Solution

Use semantic chunking, overlap, and document-aware splitting.

Definition

The needed answer is split across chunks and not retrieved together.

Solution

Use chunk overlap, parent-child retrieval, larger chunks, or hierarchical retrieval.

Definition

Irrelevant retrieved text distracts the model.

Solution

Use stricter retrieval filters, reranking, and context pruning.

Definition

Too much retrieved content overwhelms the model.

Solution

Select only the most relevant evidence and summarize where appropriate.

Definition

The answer is not present in the knowledge base.

Solution

Update the corpus or return a clear ``not found'' response.

Definition

The vector or search index is outdated.

Solution

Schedule reindexing and monitor document freshness.

Definition

Indexed documents become outdated compared with the real world.

Solution

Track document versions, owners, expiration dates, and refresh cycles.

Definition

The system retrieves an old or incorrect version of a document.

Solution

Use version-aware metadata filters and deprecate outdated content.

Definition

The retriever returns repeated or near-identical chunks.

Solution

Apply deduplication and diversity-aware retrieval.

Definition

The query and documents are embedded in a way that fails to capture meaning.

Solution

Use stronger or domain-specific embedding models and evaluate retrieval quality.

Definition

Retrieved results are semantically related but not actually useful.

Solution

Use reranking, relevance labels, and task-specific retrieval evaluation.

Definition

The retriever misunderstands the user's search intent.

Solution

Use intent detection, query rewriting, and clarification for ambiguous queries.

Definition

The system rewrites the user query incorrectly.

Solution

Evaluate rewrite quality and keep original query signals available.

Definition

Expanded terms move retrieval away from the true intent.

Solution

Limit expansion and validate expanded queries against relevance metrics.

Definition

Incorrect metadata filters exclude relevant documents.

Solution

Validate metadata quality and test filter logic.

Definition

Important metadata such as date, author, version, or product is unavailable.

Solution

Enrich documents during ingestion and enforce metadata requirements.

Definition

The correct document exists but is not included in the selected top results.

Solution

Tune top-k, retrieval scoring, reranking, and hybrid search.

Definition

The reranker fails to move the best evidence to the top.

Solution

Train or select stronger rerankers and evaluate with labeled queries.

Definition

Relevant documents are retrieved, but the model does not use them correctly.

Solution

Use answer-evidence prompts and verification checks.

Definition

The final answer is not supported by retrieved context.

Solution

Require citations, refuse unsupported claims, and validate claim-source alignment.

Definition

The model cites a source that does not support the answer.

Solution

Check each claim against its cited source before final output.

Definition

The model gives an answer without showing where it came from.

Solution

Require traceable source references for factual claims.

Definition

The answer cannot be traced back to reliable evidence.

Solution

Add source linking, evidence snippets, and audit logs.

Definition

The model retrieves the right evidence but combines it incorrectly.

Solution

Use structured synthesis prompts, chain verification, and contradiction checks.

Definition

The system fails when the answer requires multiple documents or reasoning steps.

Solution

Use iterative retrieval, graph retrieval, or query decomposition.

Definition

Search or retrieval takes too long.

Solution

Optimize indexes, cache results, reduce candidate sets, and tune infrastructure.

Definition

Similarity search fails to find the most useful content.

Solution

Use hybrid search, better embeddings, metadata filters, and reranking.

Definition

The model generates false or unsupported information.

Solution

Use grounding, retrieval, citations, uncertainty handling, and factuality evaluation.

Definition

The model invents facts, numbers, sources, citations, or events.

Solution

Require evidence and allow the model to say ``I do not know'' when information is missing.

Definition

The model sounds certain even when it is wrong.

Solution

Calibrate responses and require source-backed claims for factual answers.

Definition

The model gives answers that conflict with itself or known facts.

Solution

Use consistency checks, better context management, and verification prompts.

Definition

The model gives an unclear answer that can be interpreted multiple ways.

Solution

Ask clarifying questions or enforce structured output.

Definition

The model does not follow the user or system instruction.

Solution

Use clearer prompts, examples, schemas, and output validators.

Definition

The model receives competing instructions and follows the wrong one.

Solution

Define instruction priority and remove contradictions.

Definition

The model misunderstands what the user actually asked.

Solution

Add task clarification, examples, and intent checks.

Definition

The model forgets or ignores important information from earlier context.

Solution

Use conversation summaries, memory, retrieval, and better context selection.

Definition

The response slowly moves away from the original question or task.

Solution

Use tighter prompts, checkpoints, and validation against the user request.

Definition

The model repeats the same phrase, idea, or pattern.

Solution

Use decoding controls, repetition penalties, and response validation.

Definition

The model gives repetitive, generic, or overly similar answers.

Solution

Improve prompting, sampling settings, examples, and output diversity checks.

Definition

The model gives too much detail when a concise answer is needed.

Solution

Specify length, audience, and format constraints.

Definition

The model gives an incomplete answer.

Solution

Use coverage rubrics, checklists, and completeness validation.

Definition

The model adds unnecessary or unsupported information.

Solution

Limit scope and require evidence for added claims.

Definition

The model misunderstands which role or persona it should follow.

Solution

Use clear role instructions and periodic role reminders.

Definition

The output does not follow the requested format.

Solution

Use schemas, examples, structured output, and validators.

Definition

The model refuses a safe request or answers an unsafe request.

Solution

Improve safety policy interpretation and refusal evaluation.

Definition

The model produces harmful, unsafe, or policy-violating content.

Solution

Use safety filters, policy checks, red-team testing, and human review for high-risk cases.

Definition

The model generates offensive, abusive, hateful, or unsafe language.

Solution

Use toxicity filtering, safer training data, and moderation policies.

Definition

The model guesses instead of saying it does not know.

Solution

Instruct the model to express uncertainty and ask for missing information.

Definition

The model claims it can do something it cannot actually do.

Solution

Clearly define available tools, limits, and system capabilities.

Definition

The explanation does not match the real basis for the answer.

Solution

Use evidence-based explanations and separate hidden reasoning from user-facing justification.

Definition

The model produces different answers for the same input.

Solution

Lower temperature, use deterministic settings where possible, and validate outputs.

Definition

The model claims it used a tool or source when it did not.

Solution

Separate tool execution from response generation and log tool calls.

Definition

Unexpected harmful or incorrect behavior appears at scale.

Solution

Use staged rollout, monitoring, red teaming, and incident response.

Definition

The AI system behaves in a way that does not match human goals, rules, or expectations.

Solution

Improve instruction tuning, policy design, evaluation, and guardrails.