Back to the catalog

OKF Bundle — aeo-howto-generation-runtime

The sibling AgentCore runtime. Same deployment shape, same emit-only posture, same lazily-imported Bedrock client — several hard-won details

Open source Repository Open in the app JSON README (API)

About

# OKF Bundle — aeo-howto-generation-runtime

Open Knowledge Format bundle for this repo. Start at [service.md](/service.md).

## Sections

- [service.md](/service.md) — the repo as a concept; agent entry point
- [endpoints/](/endpoints/invocations.md) — the two HTTP surfaces
- [lib/](/lib/slot-resolution.md) — slot resolution, similarity, regeneration, the prompt, the model seam
- [business/](/business/emit-only.md) — domain concepts this repo owns
- [integrations/](/integrations/bedrock-claude.md) — outward calls
- [playbooks/](/playbooks/run-locally.md) — operational runbooks
- [briefs/](/briefs/index.md) — product ⇄ design, before the code
- [capabilities/](/capabilities/howto-article-generation.md) — what the system does, as scenarios. No status
- [qa/](/qa/howto-article-generation.md) — one checklist per capability; every requirement has a check
- [log.md](/log.md) — change history

Reserved: `index.md` and `log.md` are never concept documents.

Details

Kind
OKF bundles
Topic
Cloud & DevOps
Publisher
conexustech
Origin
okf_github
Category
dados
Version
0.1
Last push
2026-09-09T13:57:45Z
Repository state
ativo
Language
Python
Added
2026-09-09 05:03:59
Updated
2026-09-09 05:03:59
Origin id
conexusTech/aeo-howto-generation-runtime:okf/index.md

README

# aeo-howto-generation-runtime

Generates a how-to article for one shop from a shared template and that shop's
onboarding record — and measures how close it landed to every other shop's
article from the same template.

Deployed as an **AWS Bedrock AgentCore Runtime** (ARM64, Python 3.12). Called by
`aeo-backend`. Writes nothing.

```
POST /invocations   template + org context  ->  draft article + similarity + audit
GET  /ping          liveness
```

## Why this exists

Auto shops compete on price because nothing distinguishes them in the answers
customers actually see. How-to content published on the shop's own domain
signals expertise to answer engines — but only if each shop's article is
genuinely different from every other shop's. Fifty near-identical articles get
deduplicated down to one indexed page, and the other forty-nine shops pay for
something nobody is ever shown.

So the articles are **generated per shop**, not templated with the names
swapped, and the runtime records a lexical similarity score so a reviewer can
see when that is failing.

## Quick start

```bash
python -m venv .venv
.venv/Scripts/python -m pip install -r requirements.txt pytest httpx
bash scripts/gate.sh
```

The whole suite runs with **no AWS credentials, no network and no Bedrock SDK**.
That is a tested property, not a happy accident.

To serve it locally with canned articles:

```bash
HOWTO_GENERATION_USE_STUB_MODEL=true .venv/Scripts/python -m uvicorn app.howto_generation.server:app --port 8090
```

## Where things are

| | |
|---|---|
| `app/howto_generation/slots.py` | Which shop facts may be used, and which are refused |
| `app/howto_generation/similarity.py` | MinHash Jaccard over 5-word shingles |
| `app/howto_generation/regenerate.py` | Preserving human edits, and refusing when it cannot |
| `app/howto_generation/prompt.py` | The differentiation instruction |
| `app/howto_generation/runtime.py` | The pipeline |
| `okf/` | The knowledge bundle — **start at `okf/service.md`** |

## Docs

- What it does, as scenarios: `okf/capabilities/howto-article-generation.md`
- How each one is proven: `okf/qa/howto-article-generation.md`
- Running it here: `okf/playbooks/run-locally.md`

More