Retrieval-Augmented Generation (RAG) is the architectural pattern that lets the AI cite the carrier's actual claim-handling manual, the actual UW guide, the actual appetite sheet, the actual treaty wording, the actual bordereau template, and the actual actuarial-standards posture (ASOP No. 23, 38, 41, 56) - not the AI's training-data summary of what those documents might say. Without RAG, the AI hallucinates: "the carrier's UW guide permits the bind" when the actual guide requires referral; "the treaty's cat-aggregate threshold is 85%" when the actual treaty wording sets it at 78%; "ASOP No. 41 requires the actuary to disclose to the intended user" when the carrier's documented posture is more specific than the standard text. With RAG, the AI retrieves the relevant guide section, quotes it verbatim with a citation, and produces a decision grounded in the carrier's actual policy. This lesson walks the RAG architecture for insurance: embedding choice (OpenAI text-embedding-3-large vs. Voyage voyage-3-large vs. Cohere embed-v4 vs. domain-fine-tuned), chunking strategy (semantic vs. fixed-size vs. parent-document retrieval), retrieval and rerank pattern, prompt construction with retrieved context, citation enforcement, governance, and the document update cadence. Sample architecture diagram in prose form, with a worked example walking through an Atlanta adjuster's coverage-question query. The headline rule: RAG is what turns the AI from a generic generator into a tool that knows your shop's actual rules; the 2026 best-practice carrier runs RAG over every document a credentialed reviewer would cite, not training-data approximations of those documents.

The Six Document Types RAG Must Cover

Claim-handling manual. Jurisdiction-specific protocols (Texas §542 deadlines, Florida CRN response, California §790.03 framework, NY DFS Insurance Law §2601), coverage interpretation guidance per ISO and AAIS form, settlement authority levels per the delegation matrix ($250K adjuster, $1M supervisor, $5M VP Claims, $25M CCO), reserving methodology, fraud-referral procedures, recovery and subrogation discipline, file-note standards.

UW guide. Appetite class definitions per line and sub-line; rating-factor application; referral triggers (single-risk over treaty limit, cat-aggregate consumption thresholds, exposure-class restrictions); declination criteria; broker-clearance discipline; surplus-lines export-eligibility per state. The Dallas Acme Warehousing UW workflow's reason-code chain pulls from this document.

Appetite sheet. Target classes and sub-classes; explicitly excluded classes (e.g., crematories, fireworks, gun ranges, certain habitational risks, certain Tier-1 wind territory exposure); premium and exposure ranges; line-of-business preferences. Refreshed monthly or quarterly by the chief underwriting officer.

Treaty wording. Cession mechanics (quota-share percentage and ceding commission, cat-XOL layer mechanics); attachment, limit, hours clause (72-hour Named Storm, 168-hour Earthquake); exclusions; reinstatement provisions (1x or 2x at additional premium); original conditions clause; AI-driven UW representations (the 2025-2026 reinsurer additions). The carrier's reinsurance accounting director maintains this document.

Bordereau template. Required fields per cession type; reporting cadence (monthly, quarterly); formats accepted by the reinsurer; sub-broker submission requirements; reconciliation discipline. The MGA delegated-authority bordereau differs in structure from the treaty bordereau; both indexed.

Actuarial standards posture. ASOP No. 23 (data quality) - the carrier's data-quality discipline, materiality thresholds, imputation policies, exclusion rules; ASOP No. 38 (catastrophe models) - the carrier's cat-model selection, validation, secondary-uncertainty handling, model-blending; ASOP No. 41 (communications) - communication templates, sign-off chain, intended-user identification; ASOP No. 56 (modeling) - peer-review process, model-governance committee, modeling-attestation; ASOP No. 43 (reserving) and ASOP No. 36 (loss-reserve opinion) for reserve work.

The carrier's 2026 RAG corpus. A mid-size carrier's RAG corpus typically contains 800-2,500 documents totaling 8-25 million tokens of indexed content. Each document is chunked, embedded, indexed in a vector database (Pinecone, Weaviate, Chroma, or a carrier-internal stack on AWS OpenSearch or Azure Cognitive Search). Every AI invocation that needs grounded context queries the RAG corpus, retrieves the top relevant chunks, and includes them in the prompt context. The Atlanta-based adjuster team's claims-handling AI queries this corpus on every coverage question, every ROR consideration, every reserve recommendation; the Boston-based actuarial team's pricing-model AI queries it on every ASOP posture and every regulatory-citation question.

Embedding Choice

Embedding models in 2026: OpenAI text-embedding-3-large (3072 dimensions, broad domain coverage, $0.13 per 1M tokens at the API level); Voyage AI voyage-3-large (1024 or 2048 dimensions, domain-specific variants including legal and financial); Cohere embed-v4 (1024 dimensions with multimodal extension). For insurance, the Voyage legal variant or a domain-tuned embedding outperforms general embeddings by 4-8% on insurance-specific retrieval tasks (treaty-clause matching, ASOP section matching, UW-guide query, jurisdiction-specific claim-handling protocol query). Some carriers fine-tune on their own corpus (using their own claim-handling manual, UW guide, and prior-policy archive) for an additional 2-3% retrieval lift. Trade-off: general embeddings offer simpler procurement and continuous model updates; domain-specific embeddings require fine-tune maintenance and validation.

The retrieval performance question. Retrieval precision-at-5 (proportion of top-5 retrieved chunks actually relevant to the query) is the headline metric. Industry benchmarks 2026: general embedding 0.72-0.78; domain-tuned 0.80-0.86; fine-tuned on the carrier's corpus 0.82-0.88. Recall-at-5 (proportion of all relevant chunks captured in the top-5) follows a similar pattern. The few-points improvement from domain-specific embeddings translates to materially better RAG outputs because the retrieved context drives generation quality - a generic embedding that retrieves the wrong UW-guide section produces a misleading AI output even if the LLM is otherwise capable.

The 2026 best-practice carrier benchmark. Quarterly retrieval-precision benchmarks on a held-out test set of 200-500 representative queries from the carrier's actual workflow. The benchmark catches drift in retrieval quality after corpus updates, model updates, or chunking-strategy changes. The MRM analyst owns the benchmark; the chief data officer or AI governance committee reviews quarterly.

Chunking Strategy

How to split documents into chunks for retrieval. Fixed-size chunking (e.g., 512 tokens with 50-token overlap) is simple but ignores semantic boundaries - a chunk may end mid-sentence in a treaty clause, fragmenting the meaning. Semantic chunking splits at section, paragraph, or topical boundaries. Hybrid chunking combines semantic boundaries with size limits (split at section, but cap at 800 tokens; if a section exceeds the cap, split at sub-section).

For insurance documents. The claim-handling manual sections often map cleanly to natural chunks (one chunk per coverage section, one chunk per jurisdiction-specific protocol). Treaty wordings chunk per clause (Named Storm Hours Clause as one chunk, UNL definition as another, Original Conditions as another). ASOPs chunk per section per the Actuarial Standards Board's published structure. UW guides chunk per appetite class (one chunk per sub-class with the targeted exposure range and the referral triggers). The 2026 best-practice is hybrid semantic chunking with 200-800 token chunks and section-boundary respect.

Sub-chunking and parent-document retrieval. Some carriers chunk at small granularity (200 tokens) for precision-of-retrieval and then retrieve the parent-document context (the full section the chunk came from) for generation. The pattern: small chunks improve retrieval precision (the right chunk surfaces); parent-document context improves generation quality (the full section's context informs the AI's response). Pinecone, Weaviate, and LangChain libraries support this pattern with built-in parent-document retrievers.

Metadata enrichment. Every chunk carries metadata: document ID, document type (claim-handling manual / UW guide / treaty wording / etc.), section number, jurisdiction (if state-specific), effective date, last-update timestamp, document version. The metadata supports filtered retrieval - when the Atlanta adjuster queries about a Georgia coverage question, the retrieval filters to documents flagged for Georgia or for all-jurisdiction. Metadata filtering eliminates obviously-irrelevant chunks before the rerank stage.

Retrieval and Rerank

Two-stage retrieval is standard in 2026. Stage 1: vector similarity search retrieves the top-50 candidates from the corpus based on semantic similarity to the query embedding. Stage 2: a rerank model (Cohere rerank-v4, BAAI bge-reranker-v2, Voyage rerank-v2) re-orders the candidates based on query-document relevance via a cross-encoder scoring; the top-5 to top-10 reranked candidates are included in the prompt context. Rerank typically improves precision-at-5 by 8-15% over pure vector similarity.

Why rerank improves precision. The vector-similarity search is a bi-encoder pattern - the query and the document are embedded independently, then compared. Bi-encoders are fast and scalable but lose some signal because the query and document are processed in isolation. A cross-encoder reads the query and the candidate document together and produces a relevance score; cross-encoders are slower but more accurate. The two-stage pattern uses fast bi-encoder retrieval to narrow to 50 candidates, then slow cross-encoder rerank to refine to 5-10.

Query expansion. Before retrieval, expand the user's query to capture insurance-specific terminology. Example: "is the claim covered?" might expand to include "coverage analysis," "trigger of coverage," "exclusion application," "named insured verification," "endorsement schedule." Hyde (Hypothetical Document Embeddings) is the 2026 technique: generate a hypothetical answer to the query, embed the hypothetical, retrieve against the hypothetical's embedding. Hyde improves retrieval recall by 5-12% on insurance queries where the query phrasing differs from the document phrasing.

Hybrid retrieval. Some 2026 best-practice carriers combine vector retrieval with keyword search (BM25 against the same corpus) and fuse the results. The hybrid approach catches exact-match terminology (ACORD form numbers, statute citations, ASOP section numbers) that pure semantic embedding may miss. Reciprocal Rank Fusion (RRF) is the common merging strategy.

Prompt Construction With Retrieved Context

The RAG prompt includes the retrieved context explicitly. Template:

You are a [persona - e.g., senior claims examiner with AIC and 12 years of experience]. Use ONLY the following retrieved context from the carrier's [document types: claim-handling manual, UW guide, treaty wording] to answer the question. If the context does not contain the answer, say so explicitly and identify what additional information would be needed. Do NOT invent rules from training data; do NOT cite documents not in the retrieved context.

RETRIEVED CONTEXT:
[Chunk 1 from Claim-Handling Manual, Section 4.2 - Texas §542 Prompt-Pay, version 2026-Q1]
[Chunk 2 from Claim-Handling Manual, Section 4.7 - Multi-Claimant BI Escalation, version 2026-Q1]
[Chunk 3 from UW Guide, Section 8.3 - Commercial Auto Appetite, version 2026-04]

QUESTION:
[user query - e.g., "On the Dallas auto-and-GL file with three BI claimants and a §542 14-day cadence, what's the escalation path and the file-note requirement?"]

ANSWER:
Provide your answer with explicit citations to the retrieved context. Format citations as [Doc Section, version]. If the retrieved context conflicts, surface the conflict and recommend escalation.

Citation enforcement. The verification layer (the same discipline covered in the verification-workflow lesson) checks: every assertion in the AI output traces to a retrieved chunk; citations match retrieved chunks (the cited section actually was retrieved); no claims are unsupported by retrieved context. AI-generated content that does not cite, or that cites a non-existent chunk, fails verification and returns to the analyst's queue. Citation enforcement is the discipline that turns RAG from a vague-citation-system into a defensible-citation-system that survives DOI examination, bad-faith litigation discovery, and treaty-renewal scrutiny.

The conflict-surfacing requirement. When retrieved chunks conflict (the claim-handling manual says one thing, the UW guide says another), the AI must surface the conflict rather than silently picking one. The credentialed reviewer (claims supervisor or coverage counsel) resolves the conflict; the documented resolution updates the source documents to eliminate the conflict for future queries.

RAG Architecture in Prose

The end-to-end architecture has four stages.

Stage 1 - Document ingestion pipeline. Reads source documents (PDFs of claim-handling manuals, Word documents of UW guides, structured treaty-wording databases, Excel appetite sheets, scanned bordereau templates run through Hyperscience or Indico). Parses them; chunks them per the hybrid semantic strategy; embeds chunks per the chosen model; stores in vector database with metadata (document ID, section, version, last-updated, jurisdiction, document type). Ingestion runs on a cadence - nightly batch for stable documents; same-day for high-priority updates (new treaty wording, new state DOI bulletin, new ASOP revision).

Stage 2 - Query processing. User query enters the system. Query expansion (Hyde or rule-based expansion) augments the query. Vector similarity search retrieves top-50 candidates with metadata filtering (jurisdiction, document type, effective date). Rerank model re-orders to top-10. Hybrid retrieval merges vector results with keyword-search results via RRF where configured.

Stage 3 - Generation with retrieved context. Persona prompt plus CoT structure plus retrieved context plus user question plus citation-format directive feed into the LLM. The LLM generates the AI output with explicit citations to the retrieved chunks. Generation typically returns in 3-15 seconds depending on model and context size.

Stage 4 - Citation enforcement and verification. Verification layer parses citations from the output; validates that each cited chunk was in the retrieved context; runs the broader verification checks (form-edition validation, named-insured cross-check, statute-citation verification per the catch-hallucinations lesson). Output ships if all citations are valid; otherwise returns to the reviewer's queue with the verification failure noted.

Document update cadence. Source documents change continuously. UW guide updated quarterly (the chief underwriting officer signs the quarterly update). Treaty wording updates at renewal cycle (April 1 or July 1 are common renewal dates) plus mid-term endorsements. Claim-handling manual updates on regulatory change (new NAIC bulletin, new state DOI guidance) or internal-policy change. Appetite sheet updates monthly or quarterly. ASOPs update on the Actuarial Standards Board's cadence (typically multi-year revision cycles). The RAG corpus refreshes on each update. Cadence: nightly batch update for stable documents; same-day for high-priority updates. Audit trail: document version, indexed timestamp, chunked count, embedding model version - all logged so the MRM analyst can reconstruct what the AI saw at any prior point.

RAG Failure Modes and Mitigations

Failure 1 - Retrieval miss. The relevant chunk is not retrieved because the embedding similarity is too low or the query phrasing differs from the document phrasing. Mitigation: better embeddings (domain-tuned or fine-tuned), better chunking (hybrid semantic with metadata enrichment), query expansion (Hyde), rerank stage, hybrid retrieval with keyword search.

Failure 2 - Context contamination. An irrelevant chunk is retrieved and used in generation, leading the AI to produce an answer based on tangentially related content. Mitigation: rerank stage filters tangential candidates; citation validation in verification layer catches cited chunks that do not support the assertion; prompt explicitly asks the AI to ignore non-relevant context if surfaced.

Failure 3 - Hallucination despite RAG. The AI ignores the retrieved context and invents content from training data. Mitigation: the prompt explicitly instructs "use ONLY the retrieved context"; citation enforcement requires every assertion to trace to a chunk; the verification layer catches cited-content mismatches. Hallucination-despite-RAG is rare on modern frontier models but not zero.

Failure 4 - Stale corpus. A document was updated but the RAG corpus has not refreshed; the AI cites old content as current. Mitigation: refresh cadence with documented audit trail; document-version metadata on every chunk; verification layer checks document-version freshness against the source-document repository.

Failure 5 - Chunk fragmentation. The answer spans multiple chunks and the relevant context is split across retrieved chunks in a way that breaks the AI's ability to integrate. Mitigation: parent-document retrieval pattern (retrieve the small chunk for precision but include the parent section in the context); larger chunks where semantic units are large; query reformulation to seek the integrated topic.

Failure 6 - Conflicting retrieved chunks. The claim-handling manual says one thing, the UW guide says another, the appetite sheet says a third. Mitigation: the prompt explicitly requires the AI to surface conflicts; the credentialed reviewer resolves; the documented resolution updates source documents.

Actuarial Standards Posture Over RAG

The carrier's documented posture on ASOPs is one of the most-queried RAG corpora for actuarial AI workflows. Per ASOP: ASOP No. 23 (data quality) - the carrier's data-quality discipline, materiality thresholds, imputation policies, exclusion rules, validation methods. ASOP No. 38 (catastrophe models) - cat-model selection rationale (RMS Touchstone primary, Verisk AIR secondary, KCC reconciliation), validation discipline, secondary-uncertainty handling, model-blending if used. ASOP No. 41 (communications) - communication templates per artifact type (reserve opinion, rate-filing memo, ORSA narrative), sign-off chain per artifact, intended-user identification. ASOP No. 56 (modeling) - peer-review process, model-governance committee composition, modeling-attestation discipline. ASOP No. 43 (reserving) and ASOP No. 36 (loss-reserve opinion) for the reserving discipline.

AI workflows that produce actuarial outputs (model card, opinion memo, Schedule P narrative, ORSA narrative, SAO documentation) retrieve the relevant ASOP posture and cite. Without RAG, the AI hallucinates ASOP requirements or quotes generic ASOP text that misses the carrier's specific documented posture. With RAG, the AI cites the carrier's actual posture, which is the credentialed actuary's signature artifact.

The Boston actuarial team example. A pricing actuary preparing a SERFF rate-filing memo asks the AI: "What's our ASOP No. 41 posture on intended-user identification for personal-auto rate filings in Colorado?" The AI retrieves the carrier's ASOP No. 41 posture document (the specific section on rate-filing communications); pulls the chunk; cites it; produces a draft section of the memo that aligns with the carrier's posture. The actuary reviews, customizes, and signs. The Colorado DOI examiner reading the SERFF filing sees a memo that quotes the carrier's documented posture, which the actuary can defend on examination.

Key Takeaways

• RAG over six document types: claim-handling manual, UW guide, appetite sheet, treaty wording, bordereau template, actuarial standards posture (ASOP 23/38/41/56/43/36). Mid-size carrier corpus 800-2,500 documents totaling 8-25M tokens indexed in Pinecone, Weaviate, Chroma, or carrier-internal stack on AWS OpenSearch or Azure Cognitive Search.
• Embedding choice: OpenAI text-embedding-3-large (3072 dim), Voyage voyage-3-large (1024 / 2048 dim, legal variant for insurance), Cohere embed-v4 (1024 dim). Domain-tuned beats general by 4-8% on insurance retrieval; carrier-fine-tuned adds 2-3% on top.
• Hybrid semantic chunking 200-800 tokens with section-boundary respect plus metadata enrichment (jurisdiction, document type, effective date, version). Parent-document retrieval pattern: small chunks for retrieval precision, parent section for generation context.
• Two-stage retrieval: vector similarity top-50 (bi-encoder, fast) then rerank top-10 (cross-encoder, accurate). Cohere rerank-v4, BAAI bge-reranker-v2, Voyage rerank-v2 named options. Rerank improves precision-at-5 by 8-15%.
• Query expansion via Hyde (Hypothetical Document Embeddings) improves recall by 5-12%. Hybrid retrieval combining vector with keyword search via RRF catches exact-match terminology (ACORD numbers, statute citations, ASOP sections) that semantic embedding may miss.
• Prompt template: persona, "use ONLY retrieved context" directive, retrieved chunks with metadata, user question, citation-format directive. Citation enforcement via verification layer - every assertion traces to a retrieved chunk or returns to analyst queue.
• Six RAG failure modes: retrieval miss, context contamination, hallucination despite RAG, stale corpus, chunk fragmentation, conflicting retrieved chunks. Each has named mitigation; verification layer catches several at output.
• Document update cadence: nightly batch for stable documents; same-day for high-priority (treaty wording change, new NAIC bulletin, new state DOI guidance, new ASOP revision). Audit trail: document version, indexed timestamp, chunk count, embedding model version - MRM analyst can reconstruct what AI saw at any prior point.
• ASOP posture over RAG is the credentialed actuary's signature artifact. ASOP 23 data quality, 38 cat models, 41 communications, 56 modeling, 43 reserving, 36 loss-reserve opinion - the carrier's documented posture cited in rate-filing memos, reserve opinions, ORSA narratives, SAO documentation. Without RAG, hallucination; with RAG, defensible citation.