bicardinal

Multimodal retrieval engine. Drop in PDFs, Word docs, images, audio, or plain text, bicardinal extracts, summarizes, embeds, and indexes them so you can search across everything with one call.

Install

pip install bicardinal

Set your provider keys (used for OCR, vision, transcription, and summaries):

export OPENAI_API_KEY=sk-...
export MISTRAL_API_KEY=...

Deadly simple

from pathlib import Path
from bicardinal import Bicardinal

store = Bicardinal("./data")          # a home for your collections
col = store.create("docs")            # make a collection
col.init("build")                     # open it for ingestion

# Throw any file at it, pdf, docx, png, mp3, txt...
for path in Path("./my_files").glob("*"):
    col.ingest(path.name, path.read_bytes())

col.finalize() # build the index

# Search across everything
for hit in col.search("quarterly revenue growth", k=5):
    print(f"{hit.score:.3f}  {hit.filename}#{hit.chunk_index}  {hit.raw_text[:80]}")

col.close()

That's it. No pipelines to wire, no embedding code to write.

A little more

# Rank whole files by relevance
for f in col.most_similar_files("budget forecast", k=3):
    print(f.score, f.filename)

# Search within a single document
col.search_in_file("conclusion", "report.pdf", k=5)

# Reopen later, it's all on disk
col = store.open("docs")

See examples/quickstart.py for a fuller tour.

License

Apache-2.0