RAG Connector

RAG Connector is the shared contract that lets RAGauge and Pelorus Query work with the same Retrieval-Augmented Generation pipeline as a black box — without either of them knowing how that pipeline is built.

• RAGauge evaluates a RAG system: whether it retrieves the right evidence and whether its answers are grounded in it, measured against a frozen corpus and test set.
• Pelorus Query sits on top of a RAG system and serves curated, source-grounded answers (Evidence Extracts), improving them as queries repeat.

Both need the same things from the pipeline underneath: ask it a question and get ranked chunks back, read its corpus, know which embedding space and retrieval mode produced a score, and cite a chunk by an id that means the same thing to both. RAG Connector defines that once. A pipeline that implements the contract — or is wrapped by a connector that does — can be measured by RAGauge and curated by Pelorus Query, and a chunk one of them cites resolves to the same text in the other. The bundled demo corpus and its frozen chunk set exist for exactly that: both products are built against the same 605 chunk ids.

Interoperability between those two products is why this library exists, and it is why the contract is narrow. Nothing here evaluates or curates: evaluation policy belongs to RAGauge, curation behavior to Pelorus Query. Any other tool that needs to treat a RAG pipeline as a black box can build on the same contract. (RAGauge and Pelorus Query are not public yet.)

It defines:

• the retrieval contract and portable result models;
• stable chunk identity, score, and retrieval-mode semantics;
• optional capabilities such as corpus reading, query embedding, indexed chunk vectors, generation, dataset listing, run snapshots, publication, and telemetry;
• reconstructable, secret-free connector registration;
• a conformance validator and reusable connector test kit;
• an optional FastEmbed + Chroma Reference RAG that can be provisioned from a document folder, or from the 92-document demo corpus bundled with the package;
• an ingest kit (rag_connector.ingest) for hosts that build their own corpus — deliberately outside the read-only connector contract.

The bundled Reference RAG is a known-good development and learning implementation, not a production service.

Status

Pre-1.0, and used in production by its own authors. The retrieval contract, chunk identity, the registry, the validator, document loading, fingerprinting and the Reference RAG are all implemented and covered by the test suite; they are what this package is for, and they are stable in practice.

What that means for depending on it: pin a version. The core contract — RagPipeline, ChunkRecord, RetrievedChunk, retrieval modes, the registry entry point — is the part least likely to move, and a change to it would break the authors' own connectors first. The newer optional capabilities are younger and may still gain fields. Anything listed under Reserved surface is exported but unwired: nothing produces it, the validator does not check it, and it is not something to build against yet.

Reference RAG

Install the optional implementation, provision a corpus, and validate the result through the same black-box contract an external connector is held to:

python -m pip install "rag-connector[reference]"
rag-connector reference provision \
  --collection example \
  --persist-dir ./rag_connector_data/reference
rag-connector validate \
  --connector-type reference \
  --params "{\"collection_name\":\"example\",\"persist_dir\":\"./rag_connector_data/reference\"}" \
  --query "a question your documents can answer"

With no folder named, provision uses the bundled Pelorus Space corpus — 92 short documents about a fictional interplanetary travel operator, MIT-licensed like the rest of the package. Your own documents go in its place, as the first positional argument:

rag-connector reference provision ./my-documents \
  --collection example \
  --persist-dir ./rag_connector_data/reference

Add --normalized true|false to declare whether the embedding space's vectors are L2-normalized. Omit it and the space reports "not stated", which a host reads as unverifiable — deliberately distinct from an asserted false. See the embedding space descriptor.

The provision command prints the secret-free connection document needed to reopen the same instance. Supported source formats are .txt, .md, .pdf, .docx, and .rtf.

Two things to expect on a first run. The embedding model (BAAI/bge-small-en-v1.5) is downloaded on first use into fastembed's cache, so the first provision needs network and later ones do not. And on Windows, install the reference extra into a virtualenv on a short path, or enable long-path support first: onnxruntime, which fastembed depends on, ships files whose full paths exceed the 260-character MAX_PATH limit, and pip cannot unpack them under a deep directory.

Ingest kit

Loading a folder of documents, chunking it, and fingerprinting the result is not part of the connector contract — a connector is a read interface onto a system someone else already built, and nothing in that contract writes a corpus. That work lives in rag_connector.ingest, which is declared API for the other audience: a host building a corpus it owns.

from rag_connector.ingest import chunk_loaded_documents, load_documents

documents = load_documents("./my-documents")
chunks = chunk_loaded_documents(documents, chunk_size=1000, chunk_overlap=200)

Each chunk's char_start/char_end are an invariant, not a hint: document.content[char_start:char_end] == chunk.text exactly, so a stored citation resolves back to its source. They index that normalized text — the document as loaded, line endings read as \n — and not the bytes of the file, so a CRLF source resolves by the same offsets a LF one does. A chunk id is <relative path>:chunk-<n> — booking_cancellation_policy.md:chunk-0 — and subfolders are walked while dot-directories (.git, and any tool's state folder) are not. The kit imports with no optional extras installed; a parser is imported only when a file needs one.

The frozen chunk set

The bundled corpus also ships pre-chunked: 605 records at 1000/200, with the ids anything downstream cites. Read it instead of re-chunking whenever a stored artifact refers to a chunk — re-chunking is a re-derivation, and a re-derivation that lands one character differently regrounds every citation.

from rag_connector.ingest import load_bundled_chunks
from rag_connector.reference_provision import provision_reference_rag_from_chunks

chunks = load_bundled_chunks()                        # 605 ChunkRecords
connector = provision_reference_rag_from_chunks(collection_name="frozen")

The file is one JSON object per line in the shape RAGauge already writes, so both products load the same corpus, the same split, and the same ids with no adapter between them. Regenerating it is one library call — chunks_to_jsonl(build_bundled_chunks()), written to rag_connector/corpus/<name>.chunks.jsonl — and the suite fails if the committed file and the kit ever disagree. (The repository keeps that call as tools/freeze_bundled_chunks.py, which is development tooling and is not part of the installed package.)

Reproducing it

Below is everything build_bundled_chunks() applies to the bundled documents. That call re-derives the split; load_bundled_chunks() reads the frozen file instead of re-deriving it, and is what anything citing a chunk id should use. The settings are written down because a chunk set is only trustworthy if someone else can arrive at the same one.

→ →. →? →! → space | | Trailing chunk | dropped when its span lies wholly inside its predecessor | | Whitespace | each chunk is stripped, andchar_start/char_endrecord the **stripped** span | | Source identity | the corpus-relative path, forward slashes, casefolded | | Chunk id |:chunk-, dense over emitted chunks | | File discovery | recursive, sorted by relative POSIX path, dot-directories skipped,.txt .md .pdf .docx .rtf` | | Corpus layout | the 92 bundled documents, flat |

Verify a reproduction byte-for-byte. This re-derives the split from the bundled documents and digests it, and needs no optional extras:

python -c "import hashlib; from rag_connector.ingest import build_bundled_chunks, chunks_to_jsonl; b = chunks_to_jsonl(build_bundled_chunks()).encode('utf-8'); print(hashlib.sha256(b).hexdigest(), len(b), b.count(b'\n'))"

290ca86076d9b878e54bdd2d5c1a1c870c4ab20c57edd1a013983e8782ba11fc 857219 605

That is the sha256, the byte count and the line count of the shipped file itself (LF; it is pinned to LF in .gitattributes), so digesting the file rather than a re-derivation gives the same three numbers.

Two things that are not in the table, deliberately:

• The embedding model does not affect the chunk set. It decides the vectors, not the split. The bundled datasets happen to use BAAI/bge-small-en-v1.5 (384-dim, cosine), and a different model over these same 605 chunks is a different index of the same substrate.
• Everything above is version-bound. The boundary rule and the trailing-chunk rule are behaviour, not configuration, so reproducing the hash needs the same rag-connector release as well as the same parameters. A change to either moves the frozen set, which is why the drift test exists.

One caveat if you reproduce this recipe on your own corpus rather than the bundled one: text and Markdown are deterministic, but PDF, DOCX and RTF go through third-party parsers whose output can change between their own releases. A corpus of .txt/.md reproduces exactly; one full of PDFs reproduces only against a pinned parser.

Installed connectors

List every connector registered in the current environment — the bundled Reference RAG plus any independently installed connector package:

rag-connector list

A connector that does not appear here is not installed in this environment, or its package publishes no rag_connector.connectors entry point.

Development

python -m pip install -e ".[dev,reference]"
python -m pytest
python -m ruff check .

reference is in there because the Reference RAG's own tests need chromadb and fastembed. Without it the suite still runs clean — those tests skip, naming the install that turns them back on.

See Contract and the Connector author guide.

RAG Connector is licensed under the MIT License.