⚡ forge-kb

Phantom Empire Knowledge Base · VibeCodingStudio · Jetson Orin Nano 8GB · Phoenix AZ

📚 Context7 Libraries

16,165 LiteLLM snippets · 510 PydanticAI snippets
resolve-library-id + query-docs pipeline active


LiteLLMPydanticAI FastAPIFastMCP NATSgRPC RedisNginx

🤖 HuggingFace Repos

datasets/VibeCodingStudio/forge-kb
spaces/VibeCodingStudio/forge-kb


Private DatasetStatic Space

🔁 Cron Schedule

2 AM daily: provider limit scrape + model ranking recompute + NATS event
Every 30 min: Redis heartbeat for all services

cronRedis TTLNATS

🧩 Fallback Lanes

L1: Fast classifier → L2: General → L3: Reasoning → L4: Local GGUF → L5: Degrade gracefully

LiteLLM RouterQwen2.5-3B Q4_K_M

🛠 MCP Tools Active

Context7: 2 tools
HuggingFace: 14 tools
Authenticated as VibeCodingStudio

Context7HuggingFace MCP

🔒 Stack

Pi-Agent-Core RPC · LiteLLM Gateway · A2A · FastMCP · NATS JetStream · gRPC · LSP · Nginx · Cobra CLI · PydanticAI

✅ forge-kb Build Checklist

Repository

HF dataset repo created: VibeCodingStudio/forge-kb
HF Space created: VibeCodingStudio/forge-kb (this dashboard)
Scaffold generated: 305 directories across industry/domain/service-type taxonomy
GitHub repo created and pushed: VibeCodingStudio/forge-kb (private)
Clone to Jetson at /ssd/forge-kb

Context7 Ingestion

stack-manifest.yaml written (8 libraries with Context7 IDs)
ingest.py written: resolve → query-docs → write KB entry → metadata
Resolve Context7 IDs for: FastAPI, FastMCP, NATS, Redis, Nginx, gRPC
python ingestion/ingest.py --library LiteLLM
python ingestion/ingest.py --all (full stack)
Verify docs-raw/ and docs-normalized/ populated for each service-type

Model Registry

model-metadata.schema.json pushed to HF dataset registry/
model-registry-seed.json written (Groq llama-3.1-8b, Gemini 2.0 Flash, Qwen2.5-3B local)
Add providers: Cerebras, Mistral, OpenRouter, Together, Cohere
Use hub_repo_details MCP tool to confirm Jetson model fit (context window, params)
Validate all entries against schema: python -m jsonschema

Scripts (Jetson /ssd)

scrape_provider_limits.py written (2 AM daily cron)
heartbeat.py written (every 30 min)
Install cron on Jetson: crontab -e → 0 2 * * * and */30 * * * *
Add real HTML parser to scraper (selectolax or bs4)
Test: redis-cli keys "forge:heartbeat:*"
Test: nats sub models.registry.updated

HuggingFace Jobs

Create HF Job: nightly model benchmarking (off-device, keep Jetson headroom)
Create HF Job: full ingestion run on stack-manifest updates
Schedule HF Job: daily registry sync + diff report pushed back to dataset
Use hf_jobs MCP tool to submit and monitor jobs

HuggingFace Files + Datasets

README.md pushed via write_file MCP tool
model-metadata.schema.json pushed via write_file MCP tool
Upload model-registry-seed.json via write_file
Create daily scrape Parquet dataset for provider-limits time series
Create KB entries Parquet dataset for embedding + reranking pipeline
Use hub_repo_search MCP: find GGUF models for Jetson fallback lane

Dynamic Spaces (FastMCP Tools)

Wire mcp-tools/DeepSeek-OCR-experimental into FastMCP dev_tools_server
Wire ResembleAI/Chatterbox TTS into FastMCP voice tool
Wire mcp-tools/FLUX.1-Kontext-Dev into Cakeboys creative pipeline
Wire prithivMLmods/SAM3-Image-Segmentation into CakeVoid vision pipeline

Agent Stack

PydanticAI KB entry written: pydanticai-agent-mcp-sampling.md
LiteLLM config.yaml: local-primary alias + L1→L5 fallback chain
FastMCP dev_tools_server.py scaffolded with tool schemas
Pi-Agent-Core RPC wired to FastAPI ingress
NATS subjects defined: tasks.* agents.* models.registry.* forge.heartbeat.*
Cobra CLI: status, run, schedule, models sync, agents scaffold, trace, replay
Nginx reverse proxy config (FastAPI + LiteLLM, internal only)
Logfire wired to supervisor agent + all MCP calls
ALL persistent state (logs, models, Docker, cache) on /ssd not microSD
E2E smoke test: Cobra CLI run → Pi-Agent-Core → FastAPI → LiteLLM → agent → NATS result

forge-kb · Phantom Digital LLC · PhantomClaw · Phoenix AZ · 2026