Overview

Full applications built on SIE, with the pipelines, models, and evaluation results documented in each one. Every project is self-contained: clone it, run it, learn from it.

Find the best retrieval strategy for your RAG Head-to-head retrieval ablation across six conditions and eight encoder, reranker, and multi-vector models on 1,854 SEC 10-K queries, ranked by NDCG@10.

Private fine-tuned compliance RAG Domain-tuned LoRA encoder and a custom cross-encoder that reranks and prunes context in one forward pass, all served from one SIE cluster.

Self-hosted product search in 5 min A full Amazon-style product search engine running on a laptop in 5 minutes. Uses all three SIE primitives (extract, encode, score) through three SDK calls.

Build a multimodal wine recommender with OCR A demo app that pairs preference-based wine retrieval with OCR-based label detection. Shows extract, encode, and score wired into one user-facing flow.

Find SOTA embedding models by MTEB task Describe your task in plain language and search across ~14K Hugging Face embedding models, ranked by task-specific MTEB scores.

Build a multi-modal product classifier with embeddings A structured evaluation of NLI, text retrieval, image retrieval, and cross-encoder reranking on Shopify's hierarchical product taxonomy.

Swap an OCR model with one identifier change A multi-model OCR demo: recognition VLM, end-to-end document model, and zero-shot NER all driven by the same extract call. Only the model ID changes between calls.

A Stripe Link checkout with an SIE fraud-risk gate Three SIE primitives (extract, encode, score) score every order against a corpus of past fraud patterns before the Stripe PaymentIntent is created. The risk band annotates the Stripe Link payment button, in the same UI, in the same request.

More examples on GitHub

Newer gallery projects that have not been written up as docs pages yet. Each card links to the full README and runnable code in the sie repo.

A behavioural gate that catches hijacked AI agents Judges a proposed AI agent action against that agent's own learned baseline in real time, before it reaches a downstream system. Docker Compose with a self-hosted SIE, n8n, and a mock downstream.

Multi-model contract review with the OpenAI Agents SDK An agent whose every model call (triage, orchestration, vision, OCR, embeddings, rerank, entity extraction, text-to-SQL, reasoning, and a safety guardrail) is served by one SIE cluster, each step on the right catalog model.

Vision-first document RAG Retrieves and answers questions over a multi-tenant page corpus by looking at page images, including scanned drawings, with OCR kept out of the score path.

Turn difficult PDFs into Markdown Preserves tables, reading order, headings, and form labels across real financial, academic, and government PDFs with docling behind the extract call.

Rank exact primary-source passages Tests a reranker on verbatim SEC, CMS, NTSB, and Supreme Court excerpts against an SIE endpoint running Qwen3 Reranker.

Extract custom entities from primary sources Changes zero-shot GLiNER labels across financial, healthcare, rail-safety, and legal text without retraining anything.

Search licensed images with text Recomputes a six-image hard-negative ranking from full SigLIP vectors served by the encode endpoint.

Review a published flood-insurance appeal Separates FEMA's covered stone-removal scope from excluded barge, handling, disposal, and yard costs, using the bundled public FEMA appeal and policy.

Trace a restated filing figure Follows one reported figure through an original filing, corrective notice, and restatement while preserving source status, on public SEC facts.

Reproduce CMS's L1851 documentation finding Traces a published six-month requirement against a seven-month face-to-face encounter and CMS's recoupment result, on the exact CMS published example.

Reconstruct a bearing failure Turns the NTSB's three East Palestine detector readings into a cited temperature and alert sequence without adding a new causal claim.

Make a shelf gap auditable Detects one empty facing, derives its notice and shelf-label crops by geometry, then preserves OCR evidence from a CC0 supermarket shelf image.

Submit your project

We welcome community examples. To add yours:

1. Create a subdirectory in examples/ with a short name (e.g. wikipedia-search/, pdf-rag/).
2. Include a README covering what it does, how to run it, and which SIE features it uses.
3. Keep it self-contained: include requirements.txt or package.json, a docker-compose if needed, and sample data or instructions to fetch it.
4. Open a PR against main.

Projects can be anything: a search engine, a RAG pipeline, a benchmark, a migration guide, a CLI tool. If it uses SIE, it belongs here.