RAG Retrieval

Comprehensive patterns for building production RAG systems. Each category has individual rule files in rules/ loaded on-demand.

House thresholds, fusion ordering, and the latency and quality budgets we assert live in House delta below. Vendor documentation is linked, not restated (see Upstream coverage).

Quick Reference

Total: 30 rules across 9 categories

Core RAG

Fundamental patterns for retrieval, generation, and pipeline composition.

Embeddings

Embedding models, chunking strategies, and production optimization.

Contextual Retrieval

Anthropic's context-prepending technique — 67% fewer retrieval failures.

HyDE

Hypothetical Document Embeddings for bridging vocabulary gaps.

Agentic RAG

Self-correcting retrieval with LLM-driven decision making.

Multimodal RAG

Image + text retrieval with cross-modal search.

Query Decomposition

Breaking complex queries into concepts for parallel retrieval.

Reranking

Post-retrieval re-scoring for higher precision.

PGVector

Production hybrid search with PostgreSQL.

Quick Start Example

from openai import OpenAI

client = OpenAI()

async def rag_query(question: str, top_k: int = 5) -> dict:
    """Basic RAG with citations."""
    docs = await vector_db.search(question, limit=top_k)
    context = "\n\n".join([f"[{i+1}] {doc.text}" for i, doc in enumerate(docs)])

    response = await llm.chat([
        {"role": "system", "content": "Answer with inline citations [1], [2]. Use ONLY provided context."},
        {"role": "user", "content": f"Context:\n{context}\n\nQuestion: {question}"}
    ])

    return {"answer": response.content, "sources": [d.metadata['source'] for d in docs]}

Key Decisions

Upstream coverage (do not restate)

Fetch these from the source instead of restating them here. Where a row says the house subset stays, that named file carries only the tuned values and the reason, not a tutorial.

Common Mistakes

1. No citation tracking (unverifiable answers)
2. Context too large (dilutes relevance)
3. Single retrieval method (misses keyword matches)
4. Not chunking long documents (context gets lost)
5. Embedding queries differently than documents
6. No fallback path in agentic RAG (workflow hangs)
7. Infinite rewrite loops (no retry limit)
8. Using wrong similarity metric (cosine vs euclidean)
9. Not caching embeddings (recomputing unchanged content)
10. Missing image captions in multimodal RAG (limits text search)

Evaluations

See test-cases.json for 30 test cases across all categories.

There is no pass-rate, precision, recall, or MRR floor for this skill. Earlier wording here promised those numbers; they were never measured and are written down nowhere. They are left unset rather than filled in with plausible-looking values, because a threshold nobody measured is worse than an absent one: it gets cited in review as though it means something, and passes or fails changes for no defensible reason. test-cases.json is the right substrate for establishing real ones, and a baseline run over those 30 cases would produce numbers worth asserting.

The one budget this skill does assert is latency, in House delta.

Related Skills

• ork:langgraph - LangGraph workflow patterns (for agentic RAG workflows)
• ork:golden-dataset - Evaluate retrieval quality
• ork:llm-integration - Local embeddings with nomic-embed-text
• ork:multimodal-llm - Image analysis for multimodal RAG
• ork:database-patterns - Schema design for vector search
• ork:performance - Caching repeated RAG responses

Capability Details

retrieval-patterns

Keywords: retrieval, context, chunks, relevance, rag Solves:

• Retrieve relevant context for LLM
• Implement RAG pipeline with citations
• Optimize retrieval quality

hybrid-search

Keywords: hybrid, bm25, vector, fusion, rrf Solves:

• Combine keyword and semantic search
• Implement reciprocal rank fusion
• Balance precision and recall

embeddings

Keywords: embedding, text to vector, vectorize, chunk, similarity Solves:

• Convert text to vector embeddings
• Choose embedding models and dimensions
• Implement chunking strategies

contextual-retrieval

Keywords: contextual, anthropic, context-prepend, bm25 Solves:

• Prepend context to chunks for better retrieval
• Reduce retrieval failures by 67%
• Implement hybrid BM25+vector search

hyde

Keywords: hyde, hypothetical, vocabulary mismatch Solves:

• Bridge vocabulary gaps in semantic search
• Generate hypothetical documents for embedding
• Handle abstract or conceptual queries

agentic-rag

Keywords: self-rag, crag, corrective, adaptive, grading Solves:

• Build self-correcting RAG workflows
• Grade document relevance
• Implement web search fallback

multimodal-rag

Keywords: multimodal, image, clip, vision, pdf Solves:

• Build RAG with images and text
• Cross-modal search (text → image)
• Process PDFs with mixed content

query-decomposition

Keywords: decompose, multi-concept, complex query Solves:

• Break complex queries into concepts
• Parallel retrieval per concept
• Improve coverage for compound questions

reranking

Keywords: rerank, cross-encoder, precision, scoring Solves:

• Improve search precision post-retrieval
• Score relevance with cross-encoder or LLM
• Combine multiple scoring signals

pgvector-search

Keywords: pgvector, postgresql, hnsw, tsvector, hybrid Solves:

• Production hybrid search with PostgreSQL
• HNSW vs IVFFlat index selection
• SQL-based RRF fusion

House delta

Inlined rather than placed in references/: this skill carries its house knowledge in rules/ (32 files) and has never had a references/ directory, so a lone delta file there would be the only occupant.

Embedding-model choice, chunking algorithms, vector-index tuning, reranker APIs and the query-rewriting literature are vendor and upstream territory; see Upstream coverage for where each lives. What follows is only what OrchestKit adds, contradicts, or has to warn about.

Retrieval latency budget: p95 under 500ms end to end

Source: checklists/rag-quality.md:57, the only retrieval budget this skill has ever actually asserted. It covers the whole path a user waits on (embed the query, search, rerank, assemble context), not one stage in isolation. A pipeline that clears 500ms per stage and blows it in aggregate has failed this budget.

Measure at p95, not mean. Retrieval latency is dominated by tail behaviour (cold index shards, reranker queueing), and a mean hides exactly the requests users complain about.

Corpus-specific tuning does not transfer

Fusion weights, rerank depth and hybrid alpha are properties of a corpus, not of the technique. A weighting that lifts recall on prose documentation regularly hurts it on code or tabular data, so values copied between projects are noise rather than a starting point. Re-derive them per corpus against that corpus's own eval set.

Apply metadata boosting after RRF fusion, never before

Why: boost factors multiply onto the RRF score of an already-fused result set and the list is then re-sorted, using section-title match 1.5x, section-path match 1.15x, and code block on a technical query 1.2x; boosting per-method scores before fusion destroys the property RRF depends on (it fuses by rank, not by score) and lets a lone strong keyword hit outrank a document both methods agreed on, and this exact order is what produced the +6% MRR credited to boosting in rules/pgvector-metadata.md (distilled from the retired checklists/search-implementation-checklist.md and examples/examples/orchestkit-retrieval.md; no traced incident). Upstream: https://learn.microsoft.com/en-us/azure/search/index-add-scoring-profiles

Stop raising the RRF fetch multiplier past 3x

Why: measured pass rate on the reference corpus went 87.2% at 1x, 89.5% at 2x, 91.1% at 3x, 91.3% at 4x, so the fourth multiple bought 0.2 points for a third more rows scanned per method, which is why 3x is the house default hard-coded in rules/core-hybrid-search.md and rules/pgvector-hybrid-search.md; raise it only with a golden-set number that beats this curve (distilled from the retired examples/examples/orchestkit-retrieval.md; no traced incident). Upstream: https://www.elastic.co/docs/reference/elasticsearch/rest-apis/reciprocal-rank-fusion

Gate every retrieval change on the golden-set thresholds

Why: the house bar before a retrieval change ships is pass rate at least 90% on the golden query set, precision@10 at least 0.70, recall@10 at least 0.85, and MRR at least 0.65, floors set just under the reference pipeline's observed 91.6% pass rate and 0.777 overall MRR (0.695 on the hard slice) so that a real regression trips them instead of a noisy run (distilled from the retired checklists/search-implementation-checklist.md and examples/examples/orchestkit-retrieval.md; no traced incident). Upstream: ork skill golden-dataset

Assert search latency per stage, not just end to end

Why: the integration test asserts vector search under 100ms, keyword search under 50ms, and fused hybrid under 150ms separately, against an observed 415-chunk baseline of P50 15ms / P95 32ms / P99 62ms, because a single total-latency assertion stays green while one stage silently doubles, which is precisely the shape of an index that quietly stopped being used (distilled from the retired checklists/search-implementation-checklist.md and examples/examples/orchestkit-retrieval.md; no traced incident). Upstream: https://www.postgresql.org/docs/current/using-explain.html

Budget the embedding call before tuning the index

Why: in the reference measurement the query embedding was P50 8ms of a 15ms end-to-end search against 2ms for vector search and 3ms for keyword search, and it was the throughput ceiling at roughly 120 requests per second, so cache and batch embeddings first because index tuning below that ceiling moves the total by single-digit percent (distilled from the retired examples/examples/orchestkit-retrieval.md; no traced incident). Upstream: https://docs.voyageai.com/docs/embeddings