4SHADOW ANALYTIX · Field to Decision
APPENDIX
Field to Decision · Appendix

Data to LLM: The 2026 Stack, Hands-On

Not from a course. This stack runs in production right now, and the open-source side has closed the gap enough that self-hosting is a real default choice today, not a budget compromise.

The pipeline

Raw data to LLM-ready, open source at every step.

This is the working pattern behind the NewsRx production line in Story 3 and Story 5.

1

Parse

Turn PDFs, docx, and HTML into clean text. Apache Tika or Unstructured.io.

2

Chunk

Split text into retrievable pieces. Hierarchical for structured docs (3 to 5x better F1), semantic for narrative text. LangChain / LlamaIndex splitters.

3

Embed

Convert each chunk to a vector. BGE-M3 (dense, sparse, and multi-vector in one model) or Qwen3-Embedding-8B (#1 open MTEB multilingual, 32K context).

4

Serve

Run the embedding model at scale. TEI for auto-batched production serving, or raw vLLM/SGLang DIY.

5

Store + Retrieve

Index vectors for search. Qdrant or Milvus (43K+ stars, hybrid dense+sparse). Fuse with BM25 via Reciprocal Rank Fusion, then rerank with BGE-reranker-v2.

6

Generate

Feed the top reranked chunks to the LLM. Self-hosted: vLLM serving an open model (Qwen3-32B and similar).

Where process meets the model

Ontology and semantics are part of the data, not decoration on it.

Grounding lives in a knowledge graph beside Postgres running hybrid RAG and vector search. A GenAI platform in the Cortex class doesn't just draft text. It does data pulls, and the human in the loop rarely catches a bad one, because reviewers aren't taught critical thinking and a plausible number gets accepted. The bad data flows into analysis. Business, shipping, and manufacturing paths change on it. The changed outputs get written back, and the bad data is reintroduced to the very system it came from. The loop closes, and the estate starts poisoning itself.

GENAI DATA PULL REVIEWER ACCEPTS plausible, unverified ANALYSIS + DECISIONS BUSINESS, SHIPPING, MANUFACTURING PATHS CHANGE WRITE-BACK to the estate BAD DATA REINTRODUCED TO THE SYSTEM The gate that reports green is the reviewer. Story 5's lesson, on the data estate.
The loop nobody instruments closes at the write-back.

The Countermeasure

Ground retrieval in ontology and semantics so a pull carries its meaning and lineage with it. That's the knowledge graph's job beside the vector index, and it's why GraphRAG matters below. Treat "the reviewer will catch it" as an unverified gate, the same gate that reported green in 1984. Instrument the write-back path, because that's the point of irreversible consequence — the 17 instrumented write-path call sites at NewsRx are exactly this, applied to a publishing estate.

The last 12 months

Currency is a claim you have to re-earn every year.

Dense-Only Retrieval Lost

Benchmarks (BEIR, MTEB, Anthropic's own contextual retrieval work) settled it. BM25 plus dense embeddings, fused with reciprocal rank fusion, beats either alone, and a cross-encoder reranker adds another 5 to 15 points of MRR on hard queries.

12 months ago dense-only vector search was the default pattern most tutorials taught. Now hybrid search runs in 72% of production RAG systems.

Agentic RAG Went Mainstream

The biggest paradigm shift of the year. Instead of a fixed retrieve-then-generate pipeline, the model now controls retrieval itself, deciding to ask for more evidence, rewrite its own query, or stop early (Self-RAG, FLARE patterns).

12 months ago retrieval was a preprocessing step. Now it is an action the agent takes mid-reasoning — which is exactly why ungoverned data pulls became the failure mode above.

GraphRAG Hit Production Scale

Microsoft's GraphRAG (MIT-licensed) moved knowledge-graph-augmented retrieval from research papers into real, adopted production architecture.

12 months ago GraphRAG was a research pattern. Now it is a named, shipping option for teams that need entity and relationship reasoning — ontology and semantics — not just similarity search.

Open Embeddings Closed the Gap

BGE-M3 and Qwen3-Embedding-8B now rival or beat proprietary embedding APIs on MTEB, trained on massive multilingual corpora, both freely self-hostable.

12 months ago closed APIs were the safe default for embedding quality. Now a self-hosted open-weight model is a legitimate first choice.

The Amgen build, then vs. now

The build-versus-buy line moves as platforms mature.

What It Took in 2023–2024

Amgen's first GxP-validated GenAI system for FDA NDA submissions (ICH-standard CTD content, RAG-grounded, feeding Structured Content Authoring into Veeva Vault) took a 16-person build. Six Bain consultants, three data scientists, three business analysts, three medical specialist application writers, and John owning the ICH application process. Vault had no native agentic authoring then, so the RAG-to-SCA loop was architected and governed from scratch under GxP change control.

The build was custom because the platform hadn't caught up yet.

What Ships Natively, August 2026

Veeva's own Agentic Authoring application integrates natively with Vault RIM and Word, proactively drafts submissible documents, and monitors incoming data to initiate drafting. Clinical, Regulatory, and Medical AI Agents are slated for this same month, and Veeva's roadmap calls RIM+AI the standard for 2026 to 2027.

What took a 16-person custom build in 2023 is a configurable vendor feature by 2026.

What it means now

Knowing when to stop building and start configuring.

THE INSIGHT, NOT JUST THE FACT

Built today, the Amgen system likely wouldn't need a bespoke GenAI-to-SCA pipeline at all. It would configure and extend the vendor's native agents. That's not a knock on the original build. The industry converging on the same architecture two years later is validation the design was right. What it shows is that work genuinely custom-necessary one year becomes vendor commodity a couple of years later, and knowing when to stop building and start configuring is the strategist judgment call, not a technical one.

THE LONG CURVE

"I ran plasma flow calculations on a Cray-1A for DARPA in 1980. 160 megaflops, state of the art at the time. Apple's A17 Pro does over 5 teraflops, about 31,000 times faster, in something that fits in a pocket. That's the same curve as a bespoke 2023 GenAI pipeline against a 2026 native platform feature, just compressed into two years instead of forty-five."