FlashData tools and readers for LlamaIndex

Give a research agent current Google and YouTube results, or load existing video captions into a RAG pipeline with timestamps and source links. Choose the needed sources when building your application. The package supplies 17 native FunctionTool objects for 15 capabilities and a reader for the 14 data sources. Media delivery uses job tools, not a text reader.

This is an independently maintained integration. It is not an official LlamaIndex endorsement or a claim of 15 LlamaHub listings. Install the versioned package from PyPI:

python -m pip install 'llama-index-tools-flashdata==0.1.0'

Requires Python 3.11+ and llama-index-core 0.14.25–0.14.x. Get an API key from the FlashData console and provide it through your application's secret manager as FLASHDATA_API_KEY.

Installation is free; FlashData API requests use paid Credits. Check the current pricing and source-specific charges. LlamaIndex model, embedding and customer-storage charges are separate. These examples make no model or embedding calls.

Search evidence for a research agent

from llama_index.tools.flashdata import FlashDataTools, to_documents

tools = FlashDataTools(sources=["google_search"]).to_tool_list()
output = tools[0].call(query="RAG evaluation precision recall", num=5)
raw = output.raw_output  # Complete structured API response and usage.
documents = to_documents("google_search", raw)  # No second API request.
for doc in documents:
    print(doc.metadata.get("url"), doc.text)

Pass tools to your application's LlamaIndex FunctionAgent with the model you have configured. Tool schemas contain only source arguments; API key, storage credentials, source selection and download permission belong to the application. Treat returned web/caption text as untrusted evidence, not agent instructions. Google results contain search snippets and metadata, not fetched web-page text.

Caption Documents for retrieval

from llama_index.tools.flashdata import FlashDataReader

reader = FlashDataReader(source="youtube_transcript")
documents = reader.load_data(video_id="jNQXAC9IVRw", language="en")
for doc in documents:
    print(doc.metadata.get("url"), doc.metadata.get("start"), doc.text)

Each caption segment becomes a Document with its text, available start/end, video URL, languageCode and isAutoGenerated. If segments are unavailable but fullText exists, it becomes one Document without invented timestamps. Transcript defaults to automatic captions; youtube_captions defaults to uploader captions. Neither generates speech recognition, translates, or falls back to another track.

Search readers produce one JSON Document per result record, preserving optional record fields. Image URLs/dimensions are metadata, not image embeddings. Places may have no URL. The reader does not invent missing source links. Use tool raw_output for the full response, including Google related-search lists, usage and other provider envelope fields. to_documents() converts an existing result without a second paid query.

await tool.acall(...) and await reader.aload_data(...) use LlamaIndex's native thread offloading. Cancelling an await does not cancel an accepted paid job; check FlashData Jobs before resubmitting.

All capabilities

See the task guides below for inputs, outputs, a runnable example and limitations.

Media delivery lifecycle

Configure FLASHDATA_DOWNLOAD_STORAGE as a secret JSON destination. Supported providers are S3, S3-compatible, GCS, Azure, OSS, COS, BOS and TOS. Use fields from the download contract; the packaged storage.schema.json describes the credential structure. Bucket names belong in storage_url; optional folders belong in object_prefix.

from llama_index.tools.flashdata import FlashDataTools

# Only enable after authorizing this paid write to your own destination.
submit, status, result = FlashDataTools(
    sources=["youtube_download"], allow_download=True
).to_tool_list()
receipt = submit.call(video_id="jNQXAC9IVRw", download_type="audio", output_format="m4a").raw_output
job_id = receipt["id"]  # Persist before any follow-up; never resubmit to poll.
state = status.call(job_id=job_id).raw_output
# Later, read result with the same ID. Pending jobs return status, not an error.
delivery = result.call(job_id=job_id).raw_output

Download submission is disabled by default. You can use status/result tools without enabling it or supplying destination credentials. Completed result JSON expires after 24 hours; media retention depends on your storage. Supply only content you are entitled to download and use. Storage authorization is not a training license.

Failure handling

Calls are single-attempt, bounded to 120 seconds (15-second connection timeout) and 20 MiB. Failed/ambiguous POSTs are never automatically retried. A timeout can mean the paid job was accepted: check the FlashData Jobs console before another submission. Missing credits, unauthorized sources and unavailable captions are errors, not successful empty results or reasons to silently change source.

Job reads verify the ID and youtube_download source. Download status/receipt outputs exclude storage parameters. Result JSON is fetched only from the trusted HTTPS origin returned by FlashData authentication; the API key is not sent there. Redirects are not followed automatically. Raw HTTP errors and credentials are not included in FlashDataError messages.

Local verification

uv sync --frozen
uv run --frozen pytest -q
uv run --frozen python -m build
uv run --frozen twine check dist/*
uv run --frozen python scripts/verify_install.py

Tests disable network sockets and exercise real LlamaIndex tools/readers against HTTP fixtures. They do not establish live API, model, storage-write or public directory acceptance. Shared parameter and HTTP contracts come from the reviewed FlashData Haystack adapter; contract-provenance.json records exact origins and differences. No Haystack runtime dependency is needed.