Self-dogfood Demo

<!-- translations:start -->

한국어 · 中文 · 日本語 · Русский · Español · Français · Deutsch

<!-- translations:end --> This project can index itself. The self-dogfood flow proves that Tesserae can be installed, set up inside its own repository, ingest its own docs/source/tests/scripts, optionally refresh Understand Anything and Cognee, compile graph artifacts, and build the static web frontend.

The same flow doubles as a multimodal smoke test. With --with-raganything --install-raganything --run-raganything, the dogfood compile points RAG-Anything at Tesserae's own docs/ markdown plus the docs/assets/ and project-level assets/ images. That validates the multimodal pipeline against a real, project-owned non-code corpus — covering screenshots and diagrams the text-first source loaders skip — without inventing a separate fixture set.

It also exercises the self-improvement loop. Each compile re-derives mutable memory state — decay_score, access_count, confidence, and the superseded flag — into a node_memory sidecar table inside .tesserae/sqlite.db. These scalars live in the sidecar only and never in graph.json, so a fresh dogfood compile is byte-identical on the graph while the sidecar tracks decay and recurrence. Insights that recur across >= 3 distinct sessions are reinforced with a numeric confidence in (0, 1] (3 sessions → 0.5, 4 → 0.75, 5+ → 1.0, capped), written to the sidecar and surfaced by the MCP fresh_insights tool, which by default hides findings superseded by a newer near-duplicate.

Commands

From the repository root:

# Ensure the shell command is installed.
./scripts/install.sh --dir "$PWD"
export PATH="$HOME/.local/bin:$PATH"

# (optional) install the default semantic embedding backend.
pip install 'tesserae[semantic]'

# Set up this repository as an Tesserae project.
tesserae init \
  --yes \
  --name tesserae_self \
  --source README.md \
  --source docs \
  --source tesserae \
  --source tests \
  --source scripts \
  --with-understand-anything \
  --install-understand-anything \
  --understand-anything-platform codex \
  --with-raganything \
  --install-raganything \
  --raganything-parser mineru \
  --run-raganything \
  --run-cognee \
  --install-cognee

# Compile the configured sources.
tesserae compile

# Rebuild the static frontend explicitly.
tesserae export site

# Serve locally (auto-builds the site first if needed).
tesserae serve --port 8765

Open:

http://127.0.0.1:8765/

Generated workspace

The self-demo writes generated artifacts under:

.tesserae/

Key artifacts:

.tesserae/config.json
.tesserae/graph.json
.tesserae/manifest.json
.tesserae/sqlite.db          # typed graph + node_memory sidecar + live HarnessSessionsDB
.tesserae/report.md
.tesserae/competitive_report.md
.tesserae/temporal_facts.jsonl
.tesserae/graphiti_episodes.jsonl
.tesserae/markdown_projection/
.tesserae/obsidian_vault/
.tesserae/agent_harness/
.tesserae/site/
.tesserae/cognee_bundle/

The generated workspace is intentionally not committed by default. It is reproducible from the repository source with the commands above.

Latest verified run

Verified on 2026-04-27 11:11:23 KST from the Tesserae repository itself.

Integration opt-ins (Understand Anything, cognee) are interactive wizard prompts now, not CLI flags. The non-interactive equivalent below runs tesserae init --yes (integrations OFF), enables the integrations in .tesserae/config.json (the wizard writes them under the memory_backends and external_tools keys — see the integration docs for the exact keys), then refreshes each before compiling.

install command: ./scripts/install.sh --dir /Users/neo/Developer/Projects/Tesserae --skip-shell-config
setup command:   tesserae init --yes --name tesserae_self --source README.md --source docs --source tesserae --source tests --source scripts
                 # then enable Understand Anything + cognee in .tesserae/config.json and run:
                 #   tesserae integrations refresh understand-anything
                 #   tesserae integrations refresh cognee
ingest command:  tesserae compile README.md docs --changed-only
compile command: tesserae compile
site command:    tesserae export site
serve command:   tesserae serve --host 0.0.0.0 --port 56821
local URL:       http://127.0.0.1:56821/
LAN URL:         http://192.168.45.130:56821/

Final artifact counts:

nodes:               667
edges:               1020
markdown notes:      684
obsidian notes:      686
agent harness files: 14
cognee nodes:        667
cognee edges:        1020
graphiti episodes:  1020
temporal facts:      1020
site files:          index.html, nodes/index.html, sources/index.html, graph/index.html, graph.json, search-index.json, llms.txt, llms-full.txt, manifest.json, assets/style.css, assets/app.js
node pages:          687
source pages:        56

Top node types:

CodeFunction:    452
Dependency:       55
CodeClass:        54
Concept:          51
SourceFile:       47
SourceDocument:    7
CodeProject:       1

Browser verification:

loaded title: Home · tesserae_self
visible stats: 667 nodes / 1020 edges / 55 sources / 7 types
sources page: source evidence table links to per-source pages
source detail: tesserae/frontend.py shows 41 nodes, 54 related edges, type mix, node links, and edge table
search smoke: StaticSiteBuilder returned CodeClass and StaticSiteBuilder.write_site results
console: no JavaScript errors on home, sources, source detail, or graph pages
server: TCP *:56821 LISTEN, serving via --host 0.0.0.0

What this demonstrates

• Public install path works.
• tesserae shell command works.
• A repository can attach a project-local .tesserae workspace.
• Research/documentation markdown and development-code graph nodes can coexist.
• Markdown, Obsidian, frontend, Graphiti, Cognee, SQLite, report, and agent-harness projections are produced from one graph pipeline.
• The static HTML frontend can browse the project graph without a JavaScript build step.
• The self-improvement loop runs and persists: decay, access counts, recurrence confidence, and supersede flags land in the node_memory sidecar without perturbing graph.json.
• Hybrid retrieval resolves a real semantic backend when tesserae[semantic] is installed (default auto order: model2vec → sentence-transformers → hash-bucket stub); without it, embedding retrieval degrades to the non-semantic hash-bucket stub and emits a loud warning.