ndpvt-web/aiano-enhancing-information-retrieval

This skill teaches Claude to build AI-augmented annotation systems for information retrieval dataset creation, based on the AIANO methodology. The core idea: instead of annotators writing questions and judging relevance from scratch, an LLM generates candidate questions, retrieves relevant passages via embeddings, and proposes relevance scores -- then a human annotator accepts, edits, or rejects each suggestion. This human-AI loop nearly doubles annotation throughput while improving retrieval accuracy compared to manual-only annotation.

When to Use

• When the user needs to create a question-answering dataset from a document corpus (e.g., PDFs, knowledge bases, technical docs)
• When building a relevance annotation pipeline for training or evaluating a retrieval/RAG system
• When the user wants to generate synthetic QA pairs from documents and needs a human review step for quality control
• When setting up a labeling workflow where annotators judge query-passage relevance with AI pre-annotations
• When the user asks to build tooling that accelerates manual dataset curation with LLM assistance
• When evaluating or improving an existing retrieval system by creating gold-standard test sets

Key Technique

AIANO's central insight is that the annotation bottleneck in IR dataset creation is not judgment quality but annotation initiation -- starting from a blank slate is slow. By having an LLM generate the first draft (candidate questions, suggested relevant passages, initial relevance labels), annotators shift from creation mode to review-and-refine mode, which is cognitively easier and substantially faster.

The system operates on a three-stage AI-augmented loop per document chunk: (1) Question Generation -- an LLM reads a passage and proposes natural-language questions that the passage could answer; (2) Passage Retrieval -- dense embeddings index the corpus and retrieve candidate passages for each generated question, ranked by semantic similarity; (3) Relevance Annotation -- the LLM scores each query-passage pair for relevance (e.g., 0-3 scale), and these pre-annotations are presented to the human annotator for confirmation or correction.

What makes this work in practice is the confidence-gated suggestion interface: AI suggestions come with confidence indicators, so annotators can quickly accept high-confidence items and focus effort on ambiguous cases. The human retains full override authority -- they can edit questions, reject irrelevant passages, or re-score relevance. This preserves dataset quality while cutting annotation time roughly in half.

Step-by-Step Workflow

1. Ingest and chunk the document corpus. Parse source documents (PDF, HTML, markdown) into clean text. Split into passages using structure-aware chunking -- respect paragraph, section, and heading boundaries rather than fixed token windows. Target 100-300 tokens per chunk for retrieval-friendly granularity.

2. Generate dense embeddings for all passages. Use a sentence-transformer model (e.g., all-MiniLM-L6-v2 or bge-base-en-v1.5) to embed each passage. Store embeddings in a vector index (FAISS, ChromaDB, or a simple NumPy array for small corpora) alongside passage metadata (source document, chunk index, section title).

3. Generate candidate questions per passage. For each passage, prompt an LLM with the passage text and ask it to produce 2-5 natural questions that the passage answers. Instruct the LLM to vary question types: factoid, definitional, comparative, and reasoning. Include the passage context (section title, surrounding text) to improve question quality.

4. Retrieve candidate passages for each generated question. Embed each candidate question using the same embedding model. Query the vector index to retrieve the top-k (typically k=5-10) most similar passages. This step identifies which passages in the corpus are relevant to each question, including passages beyond the one that generated the question.

5. Score query-passage relevance with the LLM. For each (question, passage) pair from step 4, prompt the LLM to assign a relevance grade on a defined scale (e.g., 0=irrelevant, 1=marginally relevant, 2=relevant, 3=highly relevant). Include the grading rubric in the prompt. Attach a confidence estimate (the LLM can self-report confidence, or you can use log-probabilities if available).

6. Structure pre-annotations into a reviewable format. Organize each annotation unit as: {question, passage_text, passage_id, ai_relevance_score, ai_confidence, source_document}. Sort by confidence descending so annotators see high-confidence items first. Output as JSON, CSV, or feed into a review UI.

7. Present pre-annotations for human review. Build or use a review interface where annotators see the AI-proposed question, the candidate passage, and the suggested relevance score. Provide controls to: accept as-is, edit the question text, override the relevance score, reject the pair entirely, or flag for discussion.

8. Collect and reconcile annotations. Store human decisions alongside AI suggestions. Track acceptance rate, edit rate, and rejection rate per annotator -- these metrics indicate where the AI model is weak and where annotators disagree.

9. Export the final dataset. Produce the gold-standard dataset in a standard IR evaluation format: queries with associated relevant passage IDs and relevance grades. Common formats include TREC qrels, BEIR-compatible JSON, or HuggingFace Datasets.

10. Iterate: use annotation feedback to improve suggestions. Analyze rejection and edit patterns. Fine-tune the question generation prompt or re-rank retrieval results based on annotator corrections. Each annotation round should yield better AI suggestions for the next batch.

Concrete Examples

Example 1: Creating a QA dataset from technical documentation

User: I have 200 markdown files of API documentation. I need to create a QA
evaluation dataset with ~500 question-answer pairs to test our RAG pipeline.

Approach:
1. Parse all 200 markdown files, split by heading sections into passages
   (~100-250 tokens each). Preserve heading hierarchy as metadata.
2. Embed all passages with sentence-transformers and index in FAISS.
3. For each passage, prompt the LLM:
   "Given this API documentation passage, generate 3 realistic questions
   a developer might ask that this passage answers. Vary between
   'how-to', 'what-is', and 'troubleshooting' question types."
4. For each generated question, retrieve top-5 passages by embedding similarity.
5. Score each (question, passage) pair for relevance (0-3 scale).
6. Output a review file:

   review_batch_001.jsonl:
   {"id": "q001", "question": "How do I authenticate with OAuth2?",
    "passage": "Authentication requires an OAuth2 bearer token...",
    "passage_id": "auth-guide-chunk-04",
    "ai_relevance": 3, "ai_confidence": 0.95,
    "status": "pending_review"}

7. Annotator reviews, accepts 80%, edits 15% of questions for clarity,
   rejects 5% as duplicates or nonsensical.
8. Export final dataset as BEIR-format JSON with queries.jsonl and qrels.tsv.

Example 2: Building a relevance annotation tool for medical literature

User: We need to annotate 1000 PubMed abstracts for relevance to 50 clinical
questions. Our domain experts are expensive -- how can we speed this up?

Approach:
1. Chunk each abstract into title + full-abstract passages (medical abstracts
   are short enough to treat as single passages).
2. Embed all 1000 abstracts with a biomedical embedding model (e.g., PubMedBERT).
3. For each of the 50 clinical questions, retrieve top-20 abstracts by similarity.
4. Prompt the LLM with each (question, abstract) pair:
   "You are a medical information specialist. Rate the relevance of this
   abstract to the clinical question on a 0-3 scale:
   0 = not relevant, 1 = marginally relevant (background only),
   2 = relevant (partially answers), 3 = highly relevant (directly answers).
   Also state your confidence (low/medium/high) and a one-sentence rationale."
5. Generate a review spreadsheet sorted by question, then by AI confidence:

   | Question | Abstract ID | AI Score | Confidence | Rationale | Human Score |
   |----------|-------------|----------|------------|-----------|-------------|
   | Q12      | PMC8834521  | 3        | high       | Directly reports trial results for... | [pending] |
   | Q12      | PMC7721003  | 1        | medium     | Mentions the condition but focuses on... | [pending] |

6. Domain experts review: high-confidence items (~60%) take ~10 seconds each
   (quick accept/reject). Low-confidence items get careful reading.
7. Result: 1000 relevance judgments completed in ~40% of fully-manual time.

Example 3: Generating a synthetic QA training set with quality filtering

User: I want to generate training data for a retrieval model from our internal
wiki. No human reviewers available -- can we still use this approach?

Approach:
1. Chunk wiki pages into passages. Embed and index as usual.
2. Generate 3 questions per passage using the LLM.
3. Retrieve top-5 passages per question and score relevance.
4. Without human review, apply automated quality filters as a substitute:
   - Discard questions where AI confidence < 0.7
   - Discard questions that are too similar to each other (cosine similarity
     of question embeddings > 0.9)
   - Discard (question, passage) pairs where the top-1 retrieved passage
     is NOT the source passage (indicates the question is ambiguous)
   - Keep only pairs with AI relevance score >= 2
5. Output a silver-standard dataset. Flag it as AI-generated (not gold).
6. Use this for initial retrieval model training; create a small gold-standard
   subset with human review later for evaluation.

Note: This degraded-mode workflow trades quality for speed. The dataset will
have noise (~10-15% incorrect labels). Suitable for training, not evaluation.

Best Practices

• Do: Include the passage's surrounding context (section title, adjacent passages) when prompting for question generation. Context prevents questions that are too narrow or ambiguous.
• Do: Use a grading rubric in relevance scoring prompts. Without explicit criteria, LLM relevance scores drift across sessions and document types.
• Do: Sort review items by AI confidence descending. High-confidence items are fast to verify, building annotator momentum and catching the AI's confident mistakes early.
• Do: Track inter-annotator agreement on a sample of overlapping items. If two annotators disagree frequently on AI-suggested scores, the rubric needs refinement.
• Avoid: Generating questions from passages that are too short (< 30 tokens) or purely structural (table of contents, headers-only). These produce low-quality questions.
• Avoid: Using fixed-size chunking (e.g., every 512 tokens). Splitting mid-sentence or mid-paragraph degrades both question generation and retrieval quality. Always chunk on structural boundaries.
• Avoid: Skipping the human review step entirely for evaluation datasets. AI-generated relevance labels have systematic biases (tendency toward higher scores, inability to detect subtle factual errors) that corrupt evaluation metrics.

Error Handling

• LLM generates nonsensical or repetitive questions: Add diversity instructions to the prompt ("generate questions of different types"). Implement post-hoc deduplication by comparing question embeddings and discarding pairs with cosine similarity > 0.85.
• Retrieval returns only irrelevant passages: The embedding model may not suit the domain. Try a domain-specific model (e.g., BAAI/bge-base-en-v1.5 for general, pritamdeka/S-PubMedBert-MS-MARCO for biomedical). Also check that chunk sizes are appropriate -- very long chunks dilute embedding quality.
• AI relevance scores cluster at extremes (all 0 or all 3): The scoring prompt likely lacks calibration examples. Add 2-3 few-shot examples in the prompt showing each score level with rationale.
• Annotators disagree with AI suggestions >50% of the time: The AI suggestions are not saving time. Re-evaluate the LLM model, prompt design, or domain fit. Consider using a stronger model for suggestions or narrowing the annotation task scope.
• Corpus is too large for full embedding: Process in batches. Use approximate nearest neighbor search (FAISS IVF or HNSW) rather than brute-force. For corpora >1M passages, consider a two-stage retrieval: BM25 first-pass then re-rank with embeddings.

Limitations

• This approach assumes the source documents contain sufficient information to generate meaningful questions. Highly structured data (spreadsheets, code-only repos) may not produce good natural-language QA pairs without additional context.
• LLM-generated questions tend toward factoid and definitional types. Complex reasoning questions, multi-hop questions, and questions requiring cross-document synthesis are underrepresented and often need manual creation.
• The quality of AI pre-annotations depends heavily on the LLM's domain knowledge. For specialized domains (legal, medical, financial), general-purpose LLMs may produce plausible-sounding but technically incorrect suggestions that require expert scrutiny.
• The speed gains (roughly 2x) assume annotators trust and engage with AI suggestions. If annotators distrust the AI and re-verify everything from scratch, the pre-annotation step adds overhead rather than saving time.
• This workflow creates datasets for passage-level retrieval evaluation. It does not directly address answer extraction, multi-turn dialogue, or generative QA evaluation without adaptation.

Reference

• Paper: AIANO: Enhancing Information Retrieval with AI-Augmented Annotation (Khattab et al., 2026). Key sections: the three-stage annotation workflow (question generation, passage retrieval, relevance scoring), the within-subject user study design showing 2x speedup, and the analysis of when AI suggestions help vs. hinder annotator performance.