Magpie Search

A federated search engine — the search engine an AI agent or LLM reaches for when it needs to find something true to reason over.

Ever had your computer reboot on you, or a power outage hit mid-session? Every thread your agent was holding — gone. Now you have the tool to get it back. Never forget what your agent lost again. Magpie indexes everything your AI has ever worked through, locally, so a crash is a hiccup instead of amnesia.

What Magpie is

A normal search engine looks in one place. Magpie takes one question and fans it across everything that matters at once — the AI's entire conversation history, the files on the machine, a structured knowledge graph, a vector store, and the live web — and pulls the answer back from wherever it actually lives. Five sources, one call.

And it searches each one the right way. It can grep for an exact string or regex when you know the precise token — a file path, an error, a line of code. It can search by keyword. It can search by meaning, so it finds the thing even when the words don't match. It can do all of that at once.

Then it does the part that makes it trustworthy: it fuses everything into a single ranked answer, and every result carries a trust tier — fact > reference > lead > stale. The solid sources rise, the loose ones are marked as leads to verify, duplicates collapse, and it's all trimmed to fit so it never floods the AI's context. Ask it to go deep and it expands one question into many, reads the pages, and tells you how many independent sources agree — a full research sweep without an army of agents.

It runs entirely on the machine. No server, no account, and no telemetry unless you turn it on. The AI's transcripts and files never leave. It plugs into whatever AI is running over MCP, so the agent can reach all six sources the instant it needs them.

It is a tool for an AI — an agent or an LLM.

What's inside

At its core is a local index of the AI's transcripts: a SQLite database with two structures built side by side —

• an FTS5 full-text index (BM25 keyword ranking), and
• a vector index (sqlite-vec) of 384-dim embeddings produced locally by a small all-MiniLM-L6-v2 model.

Everything is redacted at ingest — a scrubber strips ~30 classes of secrets (keys, tokens, private keys, connection strings) before a single byte hits the index.

On top of that index sit the five search modes:

Around that sits the federation layer — the part that makes it federated:

• A provider plugin system. Five backends (transcripts, files, knowledge graph, vector, web), each returns Hit objects tagged with a trust tier.
• A fan-out: one query goes to all providers concurrently (≤8 workers), each with a 5-second timeout that fails open — a slow source contributes nothing rather than blocking the call.
• Trust-weighted RRF fusion — Reciprocal Rank Fusion where each source's rank is multiplied by its trust weight (fact ×3, reference ×2, lead ×1, stale ×0.3), damping constant 60. This is the math that merges six heterogeneous sources into one honest ranking.
• Cross-source dedup by content hash — the same fact found in three places collapses to one hit, tagged with where else it appeared (corroboration).
• A token-budget trim, so the merged set never overflows the calling AI's context.

And it exposes all of this to an AI over an MCP server — the tools it hands an agent are exactly: search, recent, session, list_sessions, stats, reindex. Note what's not in that list: nothing that writes an answer.

Why that is not RAG

RAG = Retrieval-Augmented Generation. It's a two-stage pipeline, and the defining stage is the second one: a retriever finds chunks → they're stuffed into a prompt → a language model generates the prose answer. The "G" is the whole point of the name; without a generator writing the answer, it isn't RAG.

Magpie has no G:

1. There is no generator anywhere in the search path. Nothing in Magpie composes a natural-language answer. The closest thing to a model — the cross-encoder reranker — outputs a relevance number per result and reorders the list. It scores; it never writes a sentence.
2. It stops at "here are the ranked hits." A RAG owns the prompt assembly and the model call. Magpie returns the fused, trust-ranked results and hands them back through MCP. What the AI does next — whether it even generates anything — is the AI's job, outside Magpie.
3. Its retriever is more than a RAG's retriever, not less. A textbook RAG retriever is one vector store: embed the query, top-k by cosine, done. Magpie's retrieval is six sources, five modes, trust-weighted fusion, cross-source dedup. It's a far more capable "R" — but it's still only the R.

Plug Magpie into an AI and the pair can form a RAG — Magpie is the R, the AI you bring is the G. But Magpie by itself ships only the R, and a stronger R than usual. It finds and ranks the truth; it never generates the answer.

Deep web search — research breadth without the token bill

The expensive part of "deep research" is reasoning, and the multi-agent approach pays for it N times over — one full LLM context per agent, often millions of tokens for a single question. But reasoning doesn't need to fan out; one capable model already in context can synthesize. Only the searching needs breadth — and searching the web is pure retrieval, zero LLM tokens.

magpie-search deepweb is built on that asymmetry. It fires several sub-queries at the web in parallel, fuses them by trust-weighted RRF + dedup-by-URL into one compact, token-budget-trimmed source set, optionally reads the top pages' text (still token-free), and reports how many independent domains corroborate the result — an agent-free version of the verification a research swarm pays agents to do.

So you get the breadth, page-reading, and corroboration of a multi-agent deep search, but your model only pays for a single synthesis pass over a trimmed result set.

Token cost, measured — one deep question:

That's ~2,000× fewer tokens — about 1/2000th the cost — because the searching and page-reading are pure retrieval (zero model tokens); your model only does the final synthesis pass over the trimmed, corroborated set.

# one question, several angles, read the top pages — all token-free retrieval
magpie-search deepweb "the question" --q "another angle" --q "a third angle" --thorough

The model in your loop then does one synthesis pass over the merged, corroborated set. That's the whole saving: the breadth is free, you pay only for the answer.

Install

pip install magpie-search

Or install the latest straight from source (pulls all dependencies):

pip install "git+https://github.com/xfloukiex-lab/magpie-search.git"

Optional — add the local-LLM features (the cross-encoder reranker runs on the base install; the session summarizer needs Ollama):

# 1. Install Ollama (free, runs entirely locally) — https://ollama.com/download
# 2. Pull the model magpie-search uses
ollama pull phi3.5

Python 3.10+ on Windows, macOS, and Linux.

Quickstart

magpie-search index                               # build the index (incremental)
magpie-search search "that retry backoff thing"   # keyword search
magpie-search search --mode hybrid "..."          # keyword + semantic, fused
magpie-search search --mode rerank "..."          # + cross-encoder rerank
magpie-search stats                               # sanity-check the index

Connect it to your AI (MCP)

Magpie speaks the Model Context Protocol, so any MCP-capable agent can call it. Point your client at the bundled server:

// e.g. an MCP client config
{
  "mcpServers": {
    "magpie": { "command": "magpie-search-mcp" }
  }
}

The agent then has search, recent, session, list_sessions, stats, and reindex available — federated, trust-ranked, context-budgeted.

CLI reference