dsh-ue-db — offline Unreal Engine 5.8 database for DeepSeek Harness

Ask your agent about Unreal Engine and it answers from a local corpus, without touching the network. Blueprint nodes, the C++ API, the Python API and the full Epic documentation are searchable through four tools registered in DSH.

You:    "how do I cast to another blueprint class"
Agent:  [ue_db_search] → Cast To Class (with pin types)

You:    "AActor"
Agent:  [ue_db_get] → header path, include line, full class signature

---

⏱ Read this before installing

The first download takes hours, not minutes.

The corpus is ~324 MB of documentation that is deliberately not shipped in this package (GitHub rejects files over 100 MB, and a bundled corpus would go stale). It is fetched from Epic on first use.

Epic rate-limits public documentation to roughly 4 requests per second before blocking the IP for 15–20 minutes. A complete first download therefore runs overnight:

Scope	Rough time
pyapi — Python API (plain HTTP, no Cloudflare)	~15 min
guides — 3 600 documentation pages	~45 min
bpapi — 52 000 Blueprint nodes	several hours
cppapi — C++ core modules	~2.5 h
everything	leave it running overnight

You do not have to do it in one sitting. Every step resumes: stop whenever you like, call ue_db_update again later, and it continues from where it stopped without re-downloading anything. Each step also prints its expected duration before starting, so a long run never looks like a hang.

If you only need the answer to "which node / which function", a partial download is already useful — start with pyapi, then add what you need.

---

What is inside

Source	Entries	What it gives the agent
Blueprint API	52 314 nodes	Every node with its Inputs/Outputs table: pin type, name, description
C++ API	30 841 classes	AActor, UObject, UWorld and the rest: header path, include line, full signature
Python API	11 642 classes	unreal.* signatures, parameters, properties, inheritance
Guides	3 643 pages	The Epic documentation, sorted into 36 sections: Gameplay, Animation, Niagara, UMG, PCG, Cinematics, Audio, …

98 440 entries total, about 324 MB on disk and 70 MB of index. Everything is Unreal Engine 5.8 as published on dev.epicgames.com.

The agent is told to use it

A tool description tells the model what a tool is. It does not reliably make the model reach for it — without guidance a model will happily answer an Unreal question from memory or go to the web while a complete offline corpus sits unused.

So the plugin also contributes a system-prompt section (order 210, beside the other first-party tool guidance) that says, in effect: for any Unreal Engine question consult ue_db_* first, prefer it over web search and over recollection, and only fall back to the web when the database genuinely has no answer. The section is emitted only when the tool is actually visible in the caller's scope, so an agent without the tool is not nagged about it.

You:    «как сделать двойной прыжок?»
Agent:  [ue_db_search] → JumpMaxCount …      ← not web_search, not memory

This is why you do not have to remind the agent in every conversation.

Language

The corpus is English only — that is how Epic publishes it — and the plugin does no translation at all.

That is deliberate. The agent already knows what language the user wrote in and can translate better than any fixed glossary could: a dictionary covering one language out of many was partial by construction, and it produced silently wrong answers rather than clear failures.

The contract, which lives in the tool description the model reads every turn:

Who	Responsibility
Agent	understand the question in any language, translate it to English before searching, answer in the user's language
Plugin	retrieve from an English corpus, nothing else

User (Russian):  «как создать материал»
Agent translates: "create material"            -> [ue_db_search]
Agent answers:    отвечает по-русски, используя найденное

A query containing non-Latin characters is rejected with an explicit error rather than returning an empty list — because an empty list reads as "Epic documented nothing about this", which is wrong and would send the agent searching elsewhere. The error tells the agent to translate and retry.

Install

dsh plugin --profile web add https://github.com/Abeskandr/dsh-ue-db

That is the whole install — dsh forwards to pnpm, which clones the tagged source. The corpus is not in the repository and is downloaded on first use.

Then, once:

python -m pip install cloudscraper    # needed only for downloading

Ask the agent to run ue_db_update, then restart DSH so the tools appear in the model's toolset. The first download takes hours (see the note at the top); every step resumes, so it can be done across several sittings.

Tools

Tool	Purpose
ue_db_search	Ranked search across all sources. Filter with kind=guide|pyapi|bpapi|cppapi.
ue_db_get	Full text of one entry by id from a search result.
ue_db_status	Corpus location, version, entry counts — use it to diagnose a missing base.
ue_db_update	The only tool that uses the network. Downloads and re-indexes. Resumable.

Choosing where the corpus lives

Resolution order:

1. config.knowledgePath in the plugin row — explicit path
2. DSH_UE_DB_ROOT environment variable
3. <plugin>/data when it holds an index — a local install that bundled the corpus
4. ~/.dsh/ue-db — default for a fresh registry install

- id: ue-db
  name: dsh-ue-db
  config:
    knowledgePath: D:/unreal-docs

Point it at an existing corpus and nothing is downloaded. Delete that directory and the corpus is gone; nothing is written elsewhere.

Updating

ue_db_update(what="guides")    # guides only
ue_db_update(what="bpapi")     # Blueprint API only
ue_db_update(what="cppapi")    # C++ core modules only
ue_db_update(what="cppall")    # C++ including third-party plugins — many hours
ue_db_update(limit=200)        # trial run

The crawlers pace themselves globally (EPIC_MIN_INTERVAL, default 0.35 s) and rotate the Cloudflare session on 403. Cloudflare is bypassed with cloudscraper, which imitates a browser TLS fingerprint — plain Node, curl and urllib all receive 403, which is why the network layer is Python.

Health checks

Cheap enough to run after any crawl:

python tools/check_db.py --base <dir> --full   # every entry opens
node   tools/check_queries.mjs <dir>           # 9 representative queries
node   tools/check_package.mjs                 # every tool ships in the tarball

check_db.py exists for a reason: during development two path bugs were invisible — search kept working and the index stayed valid while ue_db_get silently returned "file missing". A sampled read is not enough, hence --full.

Requirements

• DSH >= 0.2.0-rc.1
• Node 20+ (ships with DSH)
• Python 3.10+ and cloudscraper — only for ue_db_update; searching and reading are pure Node

Layout

index-ue.json      unified index (search metadata, precomputed terms)
pyapi/             Python API classes, Markdown
raw/guides/        Epic guide pages
raw/bpapi/         Blueprint API nodes
raw/cppapi/        C++ API classes
raw/*-index.json   per-source crawler indexes and resume state
tools/             Python crawlers (network layer)
lib/               Node: search engine, tools, update driver

Legal

The documentation text downloaded into pyapi/, raw/guides/, raw/bpapi/ and raw/cppapi/ is the intellectual property of Epic Games, Inc. and is subject to Epic's terms of use. This project ships only the tooling; the corpus is downloaded by the user from Epic's public documentation site for personal use.

The code — crawlers, indexer, search engine, DSH plugin — is MIT licensed.

Licence

MIT for the code. See LICENSE.