Tupshar is a knowledge store designed to fit naturally into AI and LLM workflows. This page explains what that means concretely — and what it does not mean.

Plain HTTP + JSON, No SDK Required

Every operation is a standard HTTPS request with a JSON body or query parameters. Any HTTP client — curl, ureq, fetch, requests — works without a special library. This keeps integration lightweight and avoids version coupling.

First-Class MCP Server

Tupshar ships a native MCP (Model Context Protocol) server with 16 tools covering the full API surface: create, read, update, delete, list, search (all four modes), set properties, and manage links. Claude and other MCP-capable clients can read and write your document store directly, without any glue code.

This means an LLM agent can:

Store reasoning artifacts and intermediate results as documents.
Retrieve relevant context before generating a response.
Tag and link documents to build a structured knowledge graph.
Search by keyword, tag, or property to locate what it needs.

Deterministic, Debuggable Search

Tupshar uses BM25 lexical search — term frequency, inverse document frequency, no embeddings, no neural models. Given the same document set and the same query, you get the same ranking every time. There is no "black box" score that shifts as a model is retrained.

This predictability matters in agentic pipelines: you can reason about why a document ranked high, reproduce results, and trust that the index is not silently changing behavior.

Embeddings and semantic/vector search are not implemented. They are a possible future direction. If your use case requires semantic similarity matching, Tupshar is not the right tool today.

Built for Storing and Retrieving LLM Context

Tupshar is designed around the retrieval half of RAG (Retrieval-Augmented Generation):

Store notes, summaries, tool outputs, or reference material as documents.
Tag and annotate with properties so you can filter by project, status, or any dimension you define.
Link documents to express relationships (e.g., derived_from, supersedes).
Search by keyword to pull relevant context into a prompt.

Tupshar does not reason over your documents, summarize them, or generate answers. It stores and retrieves — reliably and fast.

What Tupshar Is Not

Not a vector database. No embeddings.
Not a reasoning engine. No inference.
Not a chat interface. No conversation layer.
Not a general-purpose database. Text only, no binary, no relational joins.

It is a focused, honest tool: store text, search text, retrieve text. If that is what you need, it fits well into AI workflows.

Research Preview

Tupshar is currently a free research preview. Quotas are limited, the API surface may change, and some planned features (semantic search, cross-key sharing) are not yet available. See HTTP API for the current API surface.