Solid

Iterative deep research with an evidence score you can trust — not optimistic confidence.

Plan → search → read → score → repeat until the evidence holds up.
Follow up in the same thread. Self-hostable. Any OpenAI-compatible LLM. UI in 7 languages.

Created by Luiz Eduardo (@LuizEduPP)

Quick start · Features · How it works · API · Author · License

---

Why Solid?

Most “research agents” stop when the model feels done. Solid stops when the evidence meets explicit gates — minimum iterations, diverse sources, closed gaps, and a mandatory disconfirmation pass before high scores stick.

You get a running solidness score (0–100) backed by a visible 4-part rubric, not a black-box “confidence” number.

	Typical chat research	Solid
Stop condition	Model decides	Score + rubric gates
Source quality	Often opaque	Domains, citations, gaps tracked
High scores	Easy to inflate	Capped with open gaps; disconfirmation required
Output	One blob of text	Iterations, steps log, exportable markdown report
Deepening	Start over	Follow-up in the same session with prior context

---

Features

• Evidence-first agent loop — plans angles, searches DuckDuckGo (lite + html backends), fetches page excerpts, updates a cumulative synthesis each iteration
• Solidness panel — ring score + 4-part rubric (evidence, sources, gaps, risks) with weak / building / solid status
• Two research modes — Rigorous (100% target) and Fast (85% target); toggle from the composer (⚡ icon)
• Follow-up in the same thread — after a report, ask to deepen a point; prior score, synthesis, gaps, queries, and citations carry over
• ChatGPT-style UI — collapsible sidebar (280px / 56px rail), session history, centered empty-state composer, sticky solidness bar, glass footer
• Streaming research — live steps drawer, stop/cancel mid-run (abort signal), scroll-aware solidness pin
• Bring your own LLM — OpenAI, Ollama, LM Studio, or any /v1 compatible endpoint (default model: gpt-4o-mini, temperature 0.3)
• OpenAI-compatible API — drop-in POST /v1/chat/completions with model: "solid"
• 7 UI languages — English, Español, Português (BR/PT), Français, Deutsch, Italiano
• Local-first sessions — history (solid-history) + settings (solid-settings) in localStorage, markdown export per session

---

Quick start

Prerequisites: Node.js 20+, Yarn

git clone https://github.com/LuizEduPP/solid.git
cd solid
yarn install
yarn dev

Service	URL
Web UI	http://localhost:5173
API	http://localhost:8787
Health	http://localhost:8787/health

1. Configure your LLM

Open Settings in the sidebar:

Field	Example (local)	Example (OpenAI)
API key	(empty for local)	sk-...
Base URL	http://127.0.0.1:1234/v1	https://api.openai.com/v1
Model	your local model id	gpt-4o-mini

2. Ask a question

Type a research objective and submit. Watch iterations, rubric scores, and the final markdown report stream in.

3. Follow up (optional)

After a report finishes, stay in the same chat and ask something like “deepen point X from the report”. Solid continues from the prior synthesis, score, gaps, and sources instead of starting over.

Production build

yarn build
yarn start

Serves the built UI from the API when NODE_ENV=production. Optional .env:

cp .env.example .env
# PORT=8787
# FAVICON_CACHE_DIR=cache/favicons

---

Research modes

Thresholds are defined in src/shared.ts (MODE_THRESHOLDS).

Mode	Target score	Min. iterations	Min. unique domains	Max score Δ / iter	1st iteration cap	Disconfirm at ≥
Rigorous	100%	6	5	6	40%	70%
Fast	85%	3	3	12	55%	80%

Toggle modes with the ⚡ icon in the composer (saved in browser settings).

To reach the target score, all gates must pass: no open gaps, minimum iterations, minimum unique domains, and at least one disconfirming search round.

---

How it works

flowchart LR
  A[Your question] --> B[Planner]
  B --> C[Web search]
  C --> D[Fetch pages]
  D --> E[Analyst + rubric]
  E --> F{Gates met?}
  F -->|No| B
  F -->|Yes| G[Final report]
  G --> H[Follow-up?]
  H -->|Yes| B
  H -->|No| I[Done]

Loading

Each iteration:

1. Plan — new angle or disconfirmation query (forced when score crosses the mode threshold and no disconfirming round yet)
2. Search — DuckDuckGo via @phukon/duckduckgo-search (lite + html backends, 2.5s throttle, up to 5 retries on rate limits)
3. Read — up to 3 pages per iteration (~3,500 chars each, 8s fetch timeout); 8 search hits per query
4. Score — hybrid cumulative update (55% model / 45% rubric blend) with per-iteration caps, gap penalties, and domain caps
5. Gate — continue until all mode thresholds pass or the analyst sets should_continue: false

Scoring highlights

• Rubric: 4 × 0–25 (direct_evidence, source_diversity, gap_coverage, risk_contradiction)
• Hybrid cumulative score blended with objective signals (unique domains, citations, iterations, open gaps)
• Scores >90 capped at 90 unless ≥3 cited domains (capScoreForCitedDomains)
• Scores capped at 94 while open gaps remain
• Target score blocked while critical gaps remain, minimum iterations/domains are unmet, or disconfirmation is missing

Agent reasoning and reports follow the language of your question. The app UI is translated separately via i18n.

---

API

OpenAI-compatible routes under /v1:

Method	Path	Purpose
POST	/v1/chat/completions	Run research (streaming or not)
POST	/v1/llm/models	List models from your LLM provider
GET	/v1/models	List Solid as a model (solid)
GET	/health	Health check
GET	/favicons/:hostname	Cached favicon for a source domain

New research (streaming)

curl -N http://localhost:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "solid",
    "stream": true,
    "research_mode": "rigorous",
    "llm_api_key": "",
    "llm_base_url": "http://127.0.0.1:1234/v1",
    "llm_model": "your-model-id",
    "messages": [{"role": "user", "content": "What evidence supports X?"}]
  }'

Follow-up in the same session

Send the follow-up as the user message and include prior_context from the previous run:

curl -N http://localhost:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "solid",
    "stream": true,
    "research_mode": "rigorous",
    "llm_api_key": "",
    "llm_base_url": "http://127.0.0.1:1234/v1",
    "llm_model": "your-model-id",
    "messages": [{"role": "user", "content": "Deepen the regulatory risks section"}],
    "prior_context": {
      "rootObjective": "What evidence supports X?",
      "followUp": "Deepen the regulatory risks section",
      "cumulativeSynthesis": "...",
      "currentScore": 72.5,
      "report": "...",
      "openGaps": ["..."],
      "priorQueries": ["..."],
      "citedUrls": ["https://..."],
      "uniqueDomainCount": 4,
      "iterationCount": 3,
      "hadDisconfirmingSearch": true
    }
  }'

Request fields (optional unless noted):

Field	Default	Notes
research_mode	rigorous	rigorous or fast
target_score	mode default (100 / 85)	Override target solidness
min_score	0.01	Floor for cumulative score
temperature	0.3	LLM temperature (0–2)
prior_context	—	Resume / follow-up from prior state

Stream markers in the assistant content: @@STATUS@@ · @@SCORE@@ · @@ITER@@ · @@RUBRIC@@ · @@REPORT@@

---

Stack

Layer	Tech
Runtime	TypeScript, Node.js, ESM
API	Hono, @hono/node-server, OpenAI SDK, Zod
Agent	Custom loop, DuckDuckGo search (@phukon/duckduckgo-search), direct page fetch, ai-json-repair
UI	React 19, Vite 7, Mantine 9, react-router-dom
Markdown	react-markdown, remark-gfm, github-markdown-css
i18n	react-i18next (7 locales)

---

Project structure

public/              Static assets (logo)
src/client/          React app — UI, streaming, localStorage, locales
src/server/          Hono API — search, favicons, config
src/server/agent/    Agent loop, prompts, scoring, schemas, tests
src/shared.ts        Shared types, MODE_THRESHOLDS, PriorResearchContext, rubric helpers

---

Scripts

yarn dev         # API + Vite (ports 8787 + 5173)
yarn build       # Production client + server compile
yarn start       # Run production server
yarn typecheck   # TypeScript (client + server)
yarn test        # Agent scoring & schema tests

---

UI languages

English (default), Español, Português (Brasil), Português (Portugal), Français, Deutsch, Italiano — Settings → Language.

---

Author & attribution

Solid was created by Luiz Eduardo (@LuizEduPP).

Official repository: https://github.com/LuizEduPP/solid

If you use, fork, modify, distribute, or sell this project (including SaaS or white-label):

• Keep the LICENSE and NOTICE files in your codebase and releases.

• Credit the original author in docs, landing pages, or an About/Credits screen, for example:

  Based on Solid by Luiz Eduardo (@LuizEduPP)

Removing copyright notices from distributed copies violates the MIT License. See NOTICE for details.

---

Contributing

Issues and PRs welcome. Before submitting:

1. yarn typecheck && yarn test
2. Keep README / .env.example in sync with behavior changes
3. Match existing code style (minimal scope, no drive-by refactors)

---

License

MIT — commercial use allowed with attribution. See NOTICE.

Copyright © 2026 Luiz Eduardo.

---

If Solid helps your research workflow, star the official repo — it helps others find the original work.