An open research operating layer for traceable, book-quality knowledge work.
ENIAK is not another general agent framework. It is the kernel for a long-running research program that has to produce something a human is willing to put their name on — a chapter, a report, a literature review, an internal whitepaper, a book.
It optimises for the things most agent stacks treat as afterthoughts:
| Default in most agent stacks | ENIAK's bias |
|---|---|
| Fluent prose | Evidence traceability — every claim points back to a source |
| One-shot chat answers | Durable research state — runs are resumable, auditable, citable |
| Unchecked autonomy | Human review gates — nothing publishes without an approved state transition |
| Single-vendor lock-in | Modular providers — swap LLMs, search APIs, vector stores, publishers |
| Isolated reports | Book-quality organisation — parts, chapters, sections, cross-references |
Think of it as the operating layer between "I have a research question" and "here is a chapter I can publish to Feishu / Wiki / PDF with every claim cited and every prompt versioned."
ENIAK 不是又一个通用 Agent 框架，而是一个面向长周期研究项目的操作内核——为那些必须产出可署名内容（章节、报告、综述、白皮书、整本书）的研究流程而设计。
它把大部分 Agent 栈忽略的几件事当成一等公民：
- 证据可追溯：每一个 claim 都能回溯到 source、检索时间、页码、prompt 版本。
- 研究状态持久化：每一次 run 都可恢复、可审计、可引用。
- 人审门禁：没有经过审核态切换的内容不会被发布。
- 多供应商：LLM / 搜索 / 向量库 / 发布渠道都可换。
- 书的组织：篇 / 章 / 节 / 交叉引用，而不是孤立的报告。
定位：从"我有一个研究问题"到"这是一篇可以发到飞书 / Wiki / PDF 的章节，每个论断都有出处、每个 prompt 都有版本"之间的那一层。
ENIAK is structured as a hexagonal core with a workflow runtime, not a strict 5-layer stack. The evidence model is the domain core; the other components are ports/adapters around it.
┌──────────────────────────────┐
│ Workflow Runtime │
│ (orchestrator / runtime) │
└──────────────────────────────┘
│
┌───────────────────────────┼───────────────────────────┐
▼ ▼ ▼
┌────────────┐ ┌─────────────────┐ ┌──────────────┐
│ RADAR │ │ EVIDENCE CORE │ │ WRITER │
│ │ feeds │ │ fuels │ │
│ arXiv │──────────▶│ Source │──────────▶│ Outline │
│ S2 / OAlex │ │ Document │ │ Chapter │
│ PubMed │ │ EvidenceCard │ │ FigureSpec │
│ Web search │ │ Citation │ │ Bibliography │
│ Regulators │ │ Claim │ │ │
└────────────┘ │ Contradiction │ └──────────────┘
│ Run │ │
│ PromptTemplate │ │
│ ReviewState │ ▼
│ CostLedger │ ┌──────────────┐
└─────────────────┘ │ PUBLISHER │
│ │
│ Markdown │
│ Lark/Feishu │
│ Quarto/PDF │
└──────────────┘
Detailed design rationale lives in docs/brainstorm.md. The phase-by-phase build plan is in docs/plan.md. The full reference index (15 pinned open-source projects we learn from) is in docs/references.md and refs/MANIFEST.json.
Phase 2 — Dry-run loop live. Backend + frontend deployed. Mock radar feeds Qwen 3.5-plus extraction, citation-faithful chapter draft, Markdown export.
- UI: https://www.eniak.org — enter a topic, get an evidence-cited draft in ~30-90s.
- API: https://api.eniak.org/docs — OpenAPI playground (
POST /runs,GET /runs/{id},GET /runs/{id}/chapter.md).
research topic → mock radar (3 sources) → LLM extracts evidence cards (1 call per source)
→ LLM drafts chapter citing card IDs → claim/citation graph persisted
→ Markdown export → CostLedger entry per call
See docs/plan.md for the full roadmap (Phase 0 → 5) and
docs/DEPLOY.md for the live deployment topology.
apps/
api/ # FastAPI + SQLAlchemy 2.0 backend (Railway)
src/eniak_api/ # routers, config, app factory
migrations/ # Alembic
tests/ # pytest dry-run end-to-end
web/ # Next.js 15 (App Router) frontend (Cloudflare Pages)
src/app # routes
src/components # UI primitives
src/lib/api.ts # typed client of the FastAPI backend
packages/ # Python workspace members (uv)
eniak-evidence/ # Domain core: Source, Document, EvidenceCard, Claim, Citation,
# Contradiction, Run, PromptTemplate, ReviewState, CostLedger,
# Book, Chapter, Section
eniak-radar/ # Source monitoring (mock in Phase 2; real providers in Phase 3)
eniak-orchestrator/ # Workflow runtime (small explicit graph)
eniak-writer/ # LiteLLM client + prompts
eniak-publisher/ # Markdown (live) + Lark/Quarto/PDF (planned)
infra/
docker/ # docker-compose for local Postgres + pgvector
scripts/cf_dns_setup.sh # Idempotent DNS upsert on eniak.org
docs/
brainstorm.md # Architecture rationale
plan.md # Phased build plan
references.md # Reference project index (15 pinned repos)
DEPLOY.md # Live deployment topology + redeploy commands
refs/ # 15 pinned open-source references for local study (gitignored)
MANIFEST.json # Pinned commit SHAs (tracked)
Dockerfile # Backend image (used by Railway)
railway.json # Railway deploy config
pyproject.toml # uv workspace root
ENIAK does not reinvent. It studies and ports from 15 pinned references — see refs/MANIFEST.json for the exact commit SHAs.
| Layer | Primary references |
|---|---|
| Orchestrator | langgraph, deer-flow, open_deep_research |
| Evidence | paper-qa, graphiti |
| Writer | storm, LongWriter, quarto-cli |
| Publisher | research-claw, oapi-sdk-python |
| Prompts / Cost / Eval | langfuse, litellm, phoenix, agentevals |
| Skill banks | ljg-skills, de-anthropocentric-research-engine |
When a pattern is ported, the design doc cites the pinned SHA — e.g. "ported from storm@fb951af7".
ENIAK ships no API keys. Copy the template and fill in what you need:
cp .env.example .envRequired at minimum for the dry-run loop: an LLM gateway (OPENAI_API_KEY or any LiteLLM-compatible provider). Optional but recommended: Semantic Scholar, Tavily / Brave / Serper for web search, Lark App credentials for Feishu publishing.
Storage default is SQLite for development; Postgres + pgvector is the canonical target (see infra/docker/).
This repo is in Phase 0. Issues and design feedback are welcome, especially on:
- The hexagonal-vs-layered architecture question (see
docs/brainstorm.md) - The Python-primary vs polyglot monorepo decision (see
docs/plan.mdPhase 1) - Missing reference projects worth pinning into
refs/MANIFEST.json
PRs that add code are deferred until Phase 1 closes.
Apache License 2.0 — see also NOTICE.
Referenced projects under refs/ (gitignored) each retain their own upstream licenses.
"Evidence-native, human-gated, book-shaped."