# Local Semantic Memory Pipeline

Fully offline knowledge base for LLMs, controlled over MCP

- Role: Designer & Developer
- Period: May 2026 – present
- Status: production
- Domain: ai
- Type: case-study
- Human page: https://alicem.dev/projects/local-semantic-memory

## Summary

A document indexing pipeline that gives LLMs fast, accurate access to curated project knowledge without sending any data to external APIs.

## Case study

### Problem

Onboarding LLMs need quick access to project-specific knowledge, and none of it should leave the machine.

### Approach

A four-stage pipeline: file discovery with change detection, cleaning and translation, structured parsing, and indexing into ChromaDB with BGE-M3 embeddings. Each document gets a metadata header before embedding so vectors are identity-aware.

### Architecture

Two independent FastAPI services (pipeline and logger) controlled by a FastMCP server with 33 tools across six groups: service management, pipeline control, indexer, health, configuration and log/query.

### Result

In production since v0.4.0 (September 2026), indexing five collections and queried by several LLM clients at once. The whole system runs through natural-language tool calls, with no manual terminal work.

Note: Private production system. A live demo is available on request: contact@alicem.dev

## Highlights

- Incremental processing with crash-safe checkpoints: stop and resume at any point
- Rollback-safe production cutover with dated snapshot backups
- Config-driven: one config file drives every component, updated live over MCP
- Fully offline, no cloud dependencies

## Technologies

Python, FastAPI, FastMCP, ChromaDB, BGE-M3, Ollama, Vector Databases, Semantic Search