AI Knowledge Filler

AI-powered content production system for structured Markdown — generate, validate, and commit schema-correct files at scale.

The Problem

LLM output drifts. Wrong enum values, missing required fields, dates in the wrong format, tags: "security" instead of tags: [security]. The files look fine until something downstream breaks: a Dataview query returning nothing, a CI check failing, a search index corrupting.

Structured prompts reduce first-pass errors. They don't prevent them — model updates, provider switches, and schema edge cases still produce invalid output. Upstream reduces errors. Downstream guarantees correctness.

AKF is the production system that closes that gap.

How It Works

Prompt → LLM → Validation Engine → Error Normalizer → Retry Controller → Commit Gate → File

The LLM is the only non-deterministic component. Everything else is pure functions.

If output fails schema checks, it never touches disk — the Error Normalizer converts typed error codes into deterministic correction instructions and sends them back to the LLM.

If the same error fires twice on the same field, the pipeline aborts instead of looping — identical failure on the same field means your taxonomy has a boundary problem, not the model. More retries won't fix it.

Quick Start

pip install ai-knowledge-filler

export ANTHROPIC_API_KEY="sk-ant-..."  # or GOOGLE_API_KEY, OPENAI_API_KEY, GROQ_API_KEY

akf generate "Write a guide on Docker networking"
akf validate ./vault/

Works with Claude, GPT-4, Gemini, Groq, Ollama.

External Taxonomy Config

Your ontology lives in akf.yaml — not compiled into the tool:

# akf.yaml
schema_version: "1.0.0"
vault_path: "./vault"

enums:
  type: [concept, guide, reference, checklist, project, roadmap, template, audit]
  level: [beginner, intermediate, advanced]
  status: [draft, active, completed, archived]
  domain:
    - ai-system
    - api-design
    - devops
    - security
    - system-design

Change your taxonomy without touching code or redeploying:

akf init          # generates akf.yaml for your vault
akf validate ./   # validates all files against your config

Error Codes

Validation failures produce typed error codes, not free-form messages:

The Error Normalizer translates these into deterministic correction instructions for the retry:

E006 on field "domain" (received: "backend")
→ "The 'domain' field must be one of: [api-design, devops, security, ...]
   You used 'backend' which is not in the taxonomy. Choose the closest match."

Retry Pressure as Ontology Signal

When the same domain value triggers elevated retries across multiple generation runs, the taxonomy has a boundary problem — the model is consistently trying to say something your schema doesn't have a slot for.

The telemetry substrate (append-only JSONL) records which fields cause friction. This turns retry pressure from a failure metric into a schema health signal: evidence for refining your ontology rather than tuning your prompt.

Interfaces

CLI:

akf generate "Create a guide on API rate limiting"
akf generate "Create Docker security checklist" --model gemini
akf ask "How do I implement API rate limiting in FastAPI?" --top-k 5
akf ask "How do I implement API rate limiting in FastAPI?" --top-k 5 --no-llm
akf validate ./vault/
akf validate --file outputs/Guide.md
akf serve --port 8000        # REST API
akf serve --mcp              # MCP server

Python API:

from akf import Pipeline

pipeline = Pipeline(output="./vault/")
result = pipeline.generate("Create a guide on Docker networking")
results = pipeline.batch_generate(["Guide 1", "Guide 2", "Guide 3"])

REST API:

POST /v1/generate    →  validated file
POST /v1/enrich      →  add frontmatter to existing file
POST /v1/ask         →  RAG answer (or retrieval-only with no_llm)
POST /v1/validate    →  schema check result
POST /v1/batch       →  multiple files
GET  /v1/models      →  available providers

MCP (pip install ai-knowledge-filler[mcp]):

akf serve --mcp
# Tools: akf_generate, akf_validate, akf_enrich, akf_batch

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "akf": {
      "command": "akf",
      "args": ["serve", "--mcp"]
    }
  }
}

What Every Committed File Guarantees

• Required fields present: title, type, domain, level, status, tags, created, updated
• Valid enums: type, level, status from controlled sets
• Domain from your configured taxonomy in akf.yaml
• ISO 8601 dates with created ≤ updated
• tags as array with ≥ 3 items, title as string — no type mismatches

No file reaches disk without passing all checks.

CI Integration

# .github/workflows/validate.yml
- name: Validate docs/
  run: akf validate --path docs/

Exit code 1 on any schema error — fails the PR automatically.

Example Output

Input:

Create a guide on API rate limiting

Output (vault/API_Rate_Limiting_Strategy.md):

---
title: "API Rate Limiting Strategy"
type: guide
domain: api-design
level: intermediate
status: active
tags: [api, rate-limiting, performance, architecture]
related:
  - "[[API Design Principles]]"
  - "[[System Scalability Patterns]]"
created: 2026-03-06
updated: 2026-03-06
---

## Purpose
...

Architecture

akf/
  pipeline.py          # Pipeline — generate(), validate(), batch_generate()
  validator.py         # Validation Engine — binary VALID/INVALID, E001–E008
  validation_error.py  # ValidationError dataclass
  error_normalizer.py  # Translates errors → LLM retry instructions
  retry_controller.py  # Convergence protection — aborts on identical error hash
  commit_gate.py       # Atomic write — only VALID files reach disk
  telemetry.py         # Append-only JSONL event stream
  config.py            # Loads akf.yaml or bundled defaults
  server.py            # FastAPI REST API
  mcp_server.py        # MCP server (FastMCP)
  market_pipeline.py   # Three-stage market analysis pipeline
  defaults/
    akf.yaml           # Default taxonomy

rag/
  indexer.py           # Corpus indexer (akf index)
  retriever.py         # Semantic search layer
  copilot.py           # Retrieval + synthesis (akf ask)

cli.py                 # Entry point
llm_providers.py       # Claude / Gemini / GPT-4 / Groq / Ollama

Model Support

Tests

pytest --cov=akf --cov-report=term-missing -v

715 tests, 92% coverage, CI green on Python 3.10 / 3.11 / 3.12.

Installation

pip install ai-knowledge-filler

# With MCP support
pip install ai-knowledge-filler[mcp]

# With RAG support
pip install ai-knowledge-filler[rag]

# From source
git clone https://github.com/petro-nazarenko/ai-knowledge-filler.git
cd ai-knowledge-filler
pip install -e .

Documentation

• Architecture — module map, data flow, pipeline decisions
• CLI Reference — all commands, flags, exit codes
• User Guide — installation, configuration, troubleshooting
• REST API Threat Model — auth, endpoint exposure, limits, logging/PII
• Contributing — dev setup, quality gates, adding providers

License

MIT — free for commercial and personal use.

PyPI: https://pypi.org/project/ai-knowledge-filler