Skip to main content

Command-Line Tool

ragnet is a dotnet global tool that ingests into, and queries, a configured pipeline from the shell. It writes JSON to stdout, so its output pipes into jq and everything downstream of it.

It exists for the loop you run before writing any code: point it at a directory, ask a question, look at which chunks came back and what they scored. Answering "is my chunking sensible for this corpus" does not need an application.

Install​

dotnet tool install --global Rag.NET.Cli

Commands​

ragnet ingest <path> [--overwrite]   Ingest a file or a directory of files.
ragnet query <question> [--top-k N]  Retrieve chunks relevant to a question.

ragnet with no arguments, --help or -h prints usage and exits 0.

ragnet ingest ./docs
ragnet query "how does the retry policy back off?" --top-k 5

ingest answers with one entry per file plus a separate error list, so a directory where three files failed still reports the ones that succeeded:

{
  "Documents": [
    { "FilePath": "./docs/retrieval.md", "DocumentId": "9f2c…", "ChunksStored": 42 }
  ],
  "Errors": [
    { "FilePath": "./docs/broken.pdf", "Message": "…" }
  ]
}

query answers with the question and the ranked chunks, each carrying its score and metadata:

{
  "Query": "how does the retry policy back off?",
  "Results": [
    { "Text": "…", "Score": 0.82, "Metadata": { "source": "resilience.md" } }
  ]
}

Logs go to stderr, never stdout. Stdout is the result channel, and a log line in the middle of it would break every pipe the tool is meant to feed.

Re-ingesting the same file makes a second document​

ingest mints a fresh DocumentId per file per run rather than deriving one from the path, so running it twice over the same directory leaves two copies of everything in the store.

--overwrite does not prevent that, and it is worth being plain about why: the flag sets IngestionOptions.Overwrite, which purges that document's own prior vectors, BM25 entries and sidecar record before parsing the new content — and with a new id each run there is never a prior version of that id to purge. The flag is only meaningful to a caller that supplies a stable DocumentId, which this tool does not.

For a repeatable ingest over a changing corpus, use Rag.NET.Storage.Sqlite's content-hash record manager from your own code, where you control the id.

evaluate is not implemented​

ragnet evaluate is recognised and rejected with an explanation rather than silently missing, because "no such command" would read as a typo. The evaluators in Rag.NET.Evaluation score EvaluationSample instances that already carry a predicted answer, and no IRagEvaluator is registered by the configuration binding this tool uses. A working command needs a dataset file format and an evaluator-selection story, neither of which exists — so it was deferred rather than half-built. Use the evaluation guide and the library directly meanwhile.

Configuration​

ragnet does no pipeline wiring of its own. It binds the RagNet configuration section through Rag.NET.Hosting — the same seam Rag.NET.Mcp.Tool uses — from appsettings.json next to the working directory, or from environment variables using __ as the section separator (RagNet__ChatClient__ApiKey).

{
  "RagNet": {
    "ChatClient": {
      "Endpoint": "https://api.openai.com/v1",
      "ApiKey": "…",
      "Model": "gpt-4o-mini"
    },
    "Embeddings": {
      "Endpoint": "https://api.openai.com/v1",
      "ApiKey": "…",
      "Model": "text-embedding-3-small",
      "VectorDimensions": 1536
    },
    "VectorStore": {
      "Kind": "Qdrant",
      "Qdrant": { "Host": "localhost", "Port": 6334, "CollectionName": "ragnet" }
    }
  }
}

One OpenAI-compatible shape covers OpenAI, Azure OpenAI, OpenRouter, Ollama and LM Studio — they speak the same wire API. A local endpoint (http://localhost:11434/v1) may leave ApiKey unset; anything else requires a real key.

VectorDimensions must match what the configured embedding model actually produces — 1536 for text-embedding-3-small, 768 for nomic-embed-text — and the store has to agree with it.

The store outlives the process, or it does not​

Kind selects exactly one of InMemory, Qdrant and PgVector; only the matching block is read. InMemory is the default and needs no setup, but everything ingested is gone when the process exits — so ragnet ingest followed by ragnet query are two processes and the second finds nothing. A warning is logged at startup so this cannot pass unnoticed. Set Kind to Qdrant or PgVector for anything that has to survive between runs.

The other stores — Weaviate, Pinecone, Chroma, Azure AI Search, Redis — are not reachable from this configuration shape. Wanting one of those means hosting Rag.NET in your own application rather than driving it from this tool.

Rag.NET.Mcp.Tool is the same idea for LLM agents: a dotnet global tool over the same configuration seam, exposing the pipeline as MCP tools instead of shell commands. Configure one and you have configured the other.