The hardest part isn't the LLM layer
The hardest part of building an LLM knowledge base for a specific domain isn't wiring up the models, it's encoding how that domain thinks. A compliance analyst's knowledge base needs every obligation mapped to a specific regulation, every control to an owner, every exception to an approver. A clinical wiki needs evidence grades on every claim. A DevOps runbook system needs incident timelines and rollback steps. The query routing, intake form fields, and quality standards are all different — and none of that comes pre-configured.
Most teams spend the first two weeks not capturing knowledge but arguing about structure: what pages should exist, how routing should decompose, which source types need their own forms. By the time those questions are settled, content has already been ingested into the wrong shape and needs to be redone.
synthadoc install --template solves this in one command.
What a Domain Template Actually Provides
A domain template is not a demo or a sample dataset. It is a pre-made structural skeleton for a specific knowledge domain - everything that governs how knowledge is organized, ingested, and queried, without any of the actual content.
One install command gives you six structural components:
1. Query routing table (ROUTING.md)
The map that tells the query engine which branches of your wiki to search based on the nature of the question. A finance/investment template routes "capital structure questions" to debt financing and covenant pages, while routing "valuation questions" to financial models and comparable companies. Writing this from scratch requires domain knowledge most developers building a knowledge system do not have — and the gap shows up as poor retrieval quality until someone with that expertise intervenes.
2. Domain-tuned agent guidelines
Each template ships with CLAUDE.md, AGENTS.md, and GEMINI.md in the wiki root — the standard files that AI coding assistants (Claude Code, Gemini CLI, GitHub Copilot Agents) read automatically when active in a directory. The domain guidelines are injected into each file, so any AI tool you already use becomes domain-aware the moment it opens the wiki.
Here is what the clinical healthcare template puts in those files:
Cite clinical guidelines by issuing body, title, year, and recommendation strength (Grade A/B/C or Level 1/2/3). Note evidence quality for every clinical claim: RCT, meta-analysis, cohort study, expert consensus, or case report. Never store identifiable patient data in wiki pages.
Compare that to the compliance template:
Map every compliance obligation to the specific regulation, rule, or contractual requirement that creates it. Track controls with the control owner, testing frequency, and last test date. Distinguish between legal requirements (non-negotiable) and internal policy requirements (adjustable). Note jurisdiction and regulated entity for every obligation.
When a compliance analyst has Claude Code running in their wiki directory and asks it to help write a page or interpret a lint warning, it already knows that regulatory citations need jurisdiction-level specificity and that control owners must be named — without configuring anything.
3. Scaffold stub pages
8–12 stub pages covering the core knowledge areas of the domain — empty frames with the right titles, frontmatter, and a <!-- synthadoc:scaffold --> zone marker that separates content you write (preserved) from content the weekly scaffold job refreshes (generated). The structural skeleton maintains itself without ever overwriting your edits.
4. Blank intake forms (template-*.md)
Structured markdown forms in raw_sources/ for each document type common to the domain. A finance/investment install includes forms for deal memos, company profiles, and due diligence checklists — the same fields a professional would populate in any standard memo, now wired directly into the ingest pipeline.
5. Pre-validated seed sources (seeds.md)
Curated ingest commands for public-domain reference sources - all checked for accessibility and domain appropriateness before the template was released. For finance/investment, these include SEC investor education materials and Federal Reserve H.15 interest rate releases. For healthcare/clinical, clinical guideline registries and evidence-based medicine references. Real domain context before your proprietary content lands.
6. Scheduled maintenance jobs
A weekly lint run (Sunday 2:00 AM) and a weekly scaffold job (Sunday 3:00 AM), pre-registered at install time. Neither requires manual intervention once the wiki is running.
Template Library: 30 Domains, 9 Categories
The selection reflects domains where knowledge is both high-stakes (wrong answers have real consequences) and naturally document-heavy (the source material already exists - it is just not queryable). Finance, healthcare, and legal are the clearest examples. Technology and research are the next tier, where the cost of an engineer not knowing the current incident runbook, or a researcher duplicating a failed experiment, is concrete and measurable.
From Install to First Query
Getting a domain-template wiki to the point where it can answer real questions takes less than an hour.
Step 1: Install
synthadoc templates list # browse all 30 options
synthadoc install my-wiki --target ~/wikis --template finance/investment
synthadoc serve my-wiki
The install creates a standard wiki then applies the template: routing table, agent guidelines, stub pages, intake forms, and seeds. Staging is enabled by default — every ingested page goes to a candidates queue for review before entering the live wiki.
Step 2: Ingest seed sources and your own documents
# From seeds.md — authoritative public reference material
synthadoc ingest "https://www.investor.gov/introduction-investing" -w my-wiki
# Your own sources — batch ingest everything in raw_sources/ at once
synthadoc ingest raw_sources/ --batch -w my-wiki
Step 3: Review candidates in Obsidian
Open the Obsidian plugin and run Synthadoc: Candidates: review candidate pages... from the Command Palette. The modal lists every staged page with its slug, colour-coded confidence badge (high / medium / low), and ingest timestamp — promote what passes your bar, discard what does not.
Step 4: Run lint
synthadoc lint run -w my-wiki
Lint validates every page against the domain guidelines, checks citations, and flags orphans and contradictions. Clean draft pages are promoted to active automatically. After lint, your wiki is queryable.
What Happens After
Once the wiki is active, maintenance is driven by the web UI rather than by you remembering to check things. After every query or lint run, the engine reads its own output and pre-fills the input box with the highest-priority next action — if it found stale pages, the prompt says "Re-ingest these pages"; if re-ingest just finished, it says "Run lint to promote them"; if contradictions surfaced, it queues the resolver. You confirm and it runs. No command to remember, no separate dashboard to check.
Behind that interface, the agentic layer handles:
- Contradiction resolution — conflicting pages are identified, original sources retrieved, resolution proposed interactively
- Re-ingest on source updates — stale pages refreshed with a before/after diff for your approval
- Broken wikilink repair — dead links scanned and repaired automatically
- Citation integrity — claims audited against source documents; drift flagged for re-ingest
-
Sensitive data handling — sanitizer strips PII at ingest time;
synthadoc retractrolls back a source and all pages citing it if something slips through - Scheduled lint and scaffold — weekly jobs keep the knowledge graph healthy without intervention
The goal: after the first few days of populating your wiki, the agentic layer handles the maintenance work that would otherwise require a dedicated knowledge engineer.
Quick Reference
synthadoc templates list # browse the library
synthadoc install <name> --template <cat/domain> # install with a template
synthadoc templates sync -w <wiki> # sync after an upgrade
- Release notes (v1.3.3)
- Full template docs: template/readme in the installed package
Synthadoc is an open-source LLM knowledge compilation engine, it ingests your sources, compiles them into a queryable, citation-backed wiki, and keeps it consistent and self-growing over time. The domain template library ships in every install, no additional packages required.
Top comments (0)