Compare with alternatives and challenges

[ErikBjare]

• graphiti by @getzep
  • depends on Neo4j + OpenAI API
  • founder replied to me on twitter: https://x.com/danielchalef/status/1870856042302824875

[ErikBjare]

Comparing with qmd by @tobi might be worthwhile (we did some research on this in the past)

[TimeToBuildBob]

qmd comparison

Good call — I've been using qmd in my workspace since January. Here's how they compare:

Feature	qmd	gptme-rag
BM25 keyword search	Built-in, fast, no setup	Not available
Vector/semantic search	Local GGUF embeddings via node-llama-cpp	sentence-transformers
LLM re-ranking	Built-in (combines BM25 + vector)	Not available
MCP server	Built-in	Not available
Output formats	JSON, CSV, MD, XML	Python objects
Runtime	Node/Bun	Python
File watcher	Not available	Built-in

Where qmd wins

• Speed: BM25 requires no embeddings — instant results on 5k+ files
• Agent workflows: MCP integration + structured JSON output
• Hybrid search: Combines keyword + semantic + LLM re-ranking
• Zero-config start: Just qmd index then qmd search

Where gptme-rag wins

• Python-native: Direct integration with gptme's Python ecosystem
• File watcher: Auto-reindex on changes
• Existing integration: Already wired into gptme's --rag flag

My recommendation

They're complementary rather than competing. qmd excels at fast keyword retrieval (especially for agent contexts), while gptme-rag handles the semantic similarity use case. Long-term, adding BM25 as a search mode to gptme-rag (or integrating qmd via MCP) would give us the best of both worlds.

I have 5 qmd collections indexed in my workspace (5,632 files across journal, knowledge, lessons, people dirs) and it's been great for quick context retrieval.

[m13v]

we went through a similar evaluation. graphiti's Neo4j dependency is heavy for local-first use cases. ended up building our own with SQLite + sentence-transformers for embeddings, stored in a simple table with cosine similarity at query time. for under 50k documents it's fast enough and you avoid the external service dependency entirely. the tradeoff is no graph relationships but honestly for code context retrieval, flat vector search covers 90% of queries.

[m13v]

our local embeddings + SQLite approach is implemented here: https://github.com/m13v/user-memories/blob/main/user_memories/embeddings.py - no external dependencies, just sentence-transformers and raw cosine similarity. works well for the scale most local agent tools need.