| title | Organizational Knowledge Harness |
|---|---|
| emoji | 🕸️ |
| colorFrom | yellow |
| colorTo | gray |
| sdk | streamlit |
| sdk_version | 1.32.0 |
| app_file | app.py |
| pinned | false |
A living context engine that learns from use.
Live demo: huggingface.co/spaces/Rainvare/organizacional-knowledge-harness
Built on Harness Engineering (Böckeler, Thoughtworks 2026):
| Harness layer | Implementation |
|---|---|
| Context engineering | Extraction Agent + Graph Engine |
| Architectural constraints | NM_graph + Coherence Analyzer |
| Garbage collection | Evidence Accumulator + Proposal Queue |
| Sprint | What it does |
|---|---|
| 1 | Ingesta → Grafo → Generación con trace de nodos |
| 2 | Coherence Analyzer + Evidence Accumulator + Proposal Queue |
| 3 | Input externo — aprende de outputs de otras IAs, diferencia fuente interna/externa |
| 4 | Exportación en 4 formatos: prompt.txt, markdown, JSON, CSV |
| 5 | Input multimodal: PDF, PPTX, DOCX, imágenes, URL |
| 5+ | Smart URL parser: Behance, Figma, Dribbble con visión IA |
| 08 | Generador de prompts visuales para Midjourney, DALL-E, Sora, Runway, Kling |
The graph captures the full formal brand identity system:
| Type | What it captures |
|---|---|
brand |
Name, mission, positioning, personality |
audience |
Target segments, personas, psychographics |
value |
Core organizational values and principles |
tone |
Voice, communication style, copy rules |
restrict |
Prohibitions — what this brand never does |
example |
Reference applications, good/bad copy |
color |
Palette entries with HEX codes and usage rules |
typography |
Typefaces with name, weight, role, and hierarchy |
visual |
Logo system, layout, photography style, graphic elements |
The URL parser routes automatically by source type:
| Source | Strategy |
|---|---|
| Behance / Dribbble | Scraping + multi-image Groq vision analysis + LLM synthesis |
| Figma (public) | oEmbed metadata + thumbnail vision analysis |
| Figma (with token) | Full API — exact HEX colors, font names, component structure |
| Generic URLs | BeautifulSoup + Open Graph + JSON-LD structured data |
Paste a Behance project URL and get a typed brand graph in ~30 seconds — including color palette, typography system, logo rules, and visual language.
| Component | Technology |
|---|---|
| Frontend + Backend | Streamlit |
| LLM | Groq — llama-3.3-70b-versatile (free tier) |
| Vision | Groq — llama-3.2-11b-vision-preview |
| Graph persistence | JSON + git-native versioning |
| Hosting | Hugging Face Spaces |
git clone https://github.com/rainvare/organizational-knowledge-harness
cd organizational-knowledge-harness
pip install -r requirements.txt
# Get a free API key at console.groq.com
export GROQ_API_KEY=gsk_your_key_here
streamlit run app.pyOr enter the key directly in the sidebar of the running app.
Adapted from the Muñoz Number (UFAL, DOI: 10.5281/zenodo.18653104):
F = flagged_nodes / total_nodes
C = control_edges / total_edges
NM_graph = -ln(F / C)
> 0.5 → stable
0–0.5 → watch
≈ 0 → bifurcation (requests validation)
< 0 → unstable (blocks generation)
- Harness Engineering — Böckeler (Thoughtworks, Feb 2026)
- NM_graph — adapted from Muñoz Number / UFAL (DOI: 10.5281/zenodo.18653104)
- AI as reasoning partner — Knuth / Claude's Cycles (Stanford, Feb 2026)
- Multimodal input — MANGO (NeurIPS 2025): per-modality preprocessing prevents signal dominance
context-graph-engine → concept demo (vis.js + Claude API)
context-curator → ingestion + extraction (browser-based)
organizational-knowledge-harness → full harness platform (this repo)
Three projects, one thesis: structured context governs better than unstructured prompts.
Indira Valentina Réquiz · github.com/rainvare
Lingüística (UCV 2017) · MBA Business Intelligence (UNIR)
