Fabrik-Codek

A 7B model that knows you is worth more than a 400B that doesn't.

Fabrik-Codek is a personal cognitive architecture that runs locally with any Ollama model. It doesn't just retrieve documents — it builds a knowledge graph from how you work, measures your expertise per topic, routes tasks to the right model with the right retrieval strategy, observes whether its responses actually helped, and refines itself over time.

How It Works

1. 1. You work — Fabrik-Codek captures code changes, session transcripts, decisions, and learnings in a local datalake
2. Knowledge extraction — An 11-step pipeline extracts entities and relationships into a knowledge graph alongside a vector DB
3. Personal profiling — Analyzes your datalake to learn your domain, stack, patterns, and tooling preferences
4. Competence scoring — Measures how deep your knowledge is per topic (Expert / Competent / Novice / Unknown)
5. Adaptive routing — Classifies each query by task type and topic, selects the right model, adapts retrieval depth, and builds a 3-layer system prompt
6. Outcome tracking — Infers whether responses were useful from conversational patterns (zero friction, no manual feedback)
7. Self-correction — Adjusts retrieval parameters for underperforming task/topic combinations

Every interaction feeds back into the system. Fabrik-Codek itself makes zero outbound network requests — it only connects to Ollama and optionally Meilisearch on localhost. Model downloads are handled by Ollama's own CLI (ollama pull), not by Fabrik-Codek.

Setup

Configure as an MCP server in your openclaw.json or ~/.claude/settings.json:

CODEBLOCK0

For network access (SSE transport):

CODEBLOCK1

First Run

After installing, initialize and build the knowledge base:

CODEBLOCK2

Available MCP Tools

fabrik_ask

Ask a question to the local LLM with optional context from the knowledge base. The Task Router automatically classifies your query, selects the right model based on your competence, adapts retrieval strategy, and builds a personalized system prompt.

• - use_rag=true — vector search context
• INLINECODE4 — hybrid context (vector + graph + full-text)

Example: "How should I handle database connection pooling?"

fabrik_search

Semantic vector search across your accumulated knowledge. Returns the most relevant documents, patterns, and examples by meaning — not just keywords.

Example: "Find examples of retry logic with exponential backoff"

fabrikgraphsearch

Traverse the knowledge graph to find entities (technologies, patterns, strategies) and their relationships. Useful for understanding how concepts connect in your experience.

• - depth — how many hops to traverse (default: 2)

Example: "What technologies are related to FastAPI in my knowledge graph?"

fabrikfulltextsearch

Full-text keyword search via Meilisearch. Use this for exact keyword or phrase matching when you know the specific terms. Optional — the system works without Meilisearch installed.

Example: "Search for 'EXPLAIN ANALYZE' in my knowledge base"

fabrikgraphstats

Knowledge graph statistics: entity count, edge count, connected components, type breakdown, and relation types.

fabrik_status

System health check: Ollama availability, RAG engine, knowledge graph, full-text search, and datalake status.

Available MCP Resources

URI	Description
INLINECODE6	System component status
INLINECODE7

Knowledge graph statistics | | fabrik://config | Current configuration (sanitized) |

When to Use Each Tool

Scenario	Tool	Why
Coding question needing context	INLINECODE9 with INLINECODE10	Gets hybrid retrieval + personalized prompt
Find similar patterns or examples

fabrik_search | Semantic similarity across all knowledge | | Understand how concepts relate | fabrik_graph_search | Graph traversal shows entity relationships | | Find exact terms or phrases | fabrik_fulltext_search | BM25 keyword matching | | Check if knowledge base is healthy | fabrik_status | Component health check | | Understand knowledge distribution | fabrik_graph_stats | Entity/edge counts and types |

The Cognitive Loop

The system gets smarter the more you use it:

CODEBLOCK3

• - Personal Profile learns your domain, stack, and preferences from your datalake
• Competence Model scores expertise per topic using 4 signals (entry count, graph density, recency, outcome rate)
• Task Router classifies queries into 7 task types, detects topic, selects model, adapts retrieval
• Outcome Tracker infers response quality from conversational patterns (topic change = accepted, reformulation = rejected)
• Strategy Optimizer adjusts retrieval parameters for weak spots
• Graph Temporal Decay fades stale knowledge, reinforces recent activity
• Semantic Drift Detection alerts when an entity's context shifts between graph builds
• Context Gate decides whether to inject RAG context at all (skips for generic queries where context would be noise)
• Relevance Filter drops retrieved chunks with low query-text token overlap, preventing domain-specific knowledge from contaminating generic answers

Requirements

• - Fabrik-Codek installed from source (git clone + pip install -e ".[dev]")
• Ollama running locally with any model (e.g., ollama pull qwen2.5-coder:7b)
• Optional: Meilisearch for full-text search (system works without it)

Note on installation: Fabrik-Codek is an instruction-only skill — there is no automated installer. You install it manually from the GitHub repository via git clone + pip install -e ".[dev]". This lets you audit the full source code before installing. The skill itself contains documentation and MCP server configuration, not executable code.

Security & Privacy

No external network calls

Fabrik-Codek makes zero outbound network requests. It connects only to services running on your own machine:

• - Ollama at localhost:11434 — your locally running LLM server (for inference and embeddings)
• Meilisearch at localhost:7700 (optional) — your locally running search engine

No telemetry, no analytics, no phone-home. Verify in the source: grep -r "requests\.\|httpx\.\|urllib" src/ — all HTTP calls target localhost only. The only network activity that occurs during setup is ollama pull, which is Ollama's own CLI downloading models from ollama.ai/library — Fabrik-Codek does not initiate or control these downloads.

What fabrik init does

INLINECODE27 performs these local-only operations:

1. 1. Checks Python version (>= 3.11)
2. Detects if Ollama is running at INLINECODE28
3. Creates a .env config file in the current directory
4. Creates local data directories (./data/embeddings/, ./data/graphdb/, ./data/profile/)
5. Pulls Ollama models via ollama pull — models are downloaded by Ollama itself from ollama.ai/library, not by Fabrik-Codek

Fabrik-Codek does not download any files from any server. Model downloads are handled entirely by Ollama's own CLI.

Data access scope

Reads (all local, all opt-in, never automatic):

Path	What	When	Why
INLINECODE34	Session transcript JSONL files (already on disk from Claude Code)	Only when you explicitly run fabrik learn process or INLINECODE36	Extracts entities and reasoning patterns to build the knowledge graph. This path is NOT in configPaths because Fabrik-Codek does not write to it — it is read-only and user-initiated.
INLINECODE38 or INLINECODE39

Your datalake (training pairs, captures, metadata) | During graph build, rag index, profile build, competence build | Source data for building the knowledge base and personal profile |

Writes (all local):

Path	What
INLINECODE44	LanceDB vector index
INLINECODE45

NetworkX knowledge graph (JSON) |
| ./data/profile/ | Personal profile, competence map, strategy overrides (JSON) |
| ./data/01-raw/outcomes/ | Outcome tracking records (JSONL) |

All paths are declared in the skill metadata configPaths. The skill never writes outside these directories.

Network transport

• - Default: stdio — no network listener, no ports opened, no exposure
• Optional: sse — starts an HTTP server bound to 127.0.0.1:8421 by default (localhost only, not reachable from other machines)
• If you change the SSE bind address to 0.0.0.0, your indexed data would be accessible over the network. Do not do this without proper firewall/ACL rules

Session transcript privacy

The fabrik learn command reads Claude Code session transcripts, which may contain sensitive data (code, credentials, conversation history). This command is opt-in — you must run it manually. It does not run in the background or on a schedule unless you explicitly configure fabrik learn watch. Review what's in your ~/.claude/projects/ before indexing.

Source verification

Fully open source at github.com/ikchain/Fabrik-Codek (MIT license). Clone the repo and audit before installing.