Advanced search controls
Metadata filters, semantic ratio tuning, boosts and chunk relations, all through one query API. The lever you need is a request parameter, not a fork of the pipeline.
Now self-tuningcaptain_eval
Retrieval your agents can trust.
The complete file search platform. Synced with cloud storage, fast and accurate, and optimized for complex retrieval.
Backed by
One platform from raw files
to ranked answers.
Powerful ingestion and self-tuning retrieval,
behind one query API, ready for production scale.
01Search
Search
Sub-second queries through a deterministic API, with filters, rerank and candidate depth all yours to configure.
02Iterate03Process04Engine05Sync
01Search
Search
Sub-second queries through a deterministic API, with filters, rerank and candidate depth all yours to configure.
01Search02Iterate03Process04Engine05Sync
Sync files and self-improve search as one platform
Every layer above ships together, on one index. Nothing to wire and nothing to keep aligned.
Files in.
Answers out.
How it works
Add Captain to your coding agent.
MCP source codeclaude mcp add --transport http captain https://mcp.captain.dev/mcp \ --header "Authorization: Bearer cap_..."
Search fails at the parser,
not at the embedding.
Parsing
A DOCX or a scanned PDF is not text yet. Whatever the parser drops, no search can find.
DOCX
XML fragments in a zip, with tables that only exist as formatting. Captain's parser was tuned on the ugliest DOCX files our customers could find.
PDF
Clean PDFs parse anywhere. Captain also reads the rotated scan and the handwritten margin note.
Video + audio
Transcribed, described and indexed at ingest. The chart on slide 12 is a query, not a scrubbing session.
Your data is
the benchmark.
Benchmarks
A public benchmark tells you how a system ran on someone else's corpus. Run and Captain builds hundreds of Q&A evals from your own files, scores retrieval against them, and tunes until it hits your target.
Coverage: queries answered
Captain100%
Gemini File Search~66%
When evaluated against other leading file-search APIs on production queries, alternatives fail to answer about one in three questions.
+24%
Higher accuracyvs leading alternatives · BEIR nDCG@10
108×
Faster indexing18 hrs → 10 min
7.5×
Faster search1,500ms → 100ms
Observability01
Usage insights
See what people search for, and exactly where answers fall short. The feedback loop makes retrieval better over time.
Observability02
Knowledge gap discovery
See the questions your data has yet to answer, surfaced automatically, so you know exactly where to close the gaps.
Observability03
Retrieval attack detection
Catch black-hole vector attacks (arXiv:2604.05480) before they reach production, a failure mode most DIY search layers never see.
The accuracy is not a model.
It is one index, an eval loop,
and parsing that keeps every word.
Why captain
Multimodal & video
One index across text, documents and video. Ask where the Q3 margin chart appears and get the file, the page and the timestamp.
Cross-cloud indexing
Index straight from object storage and cloud drives. One platform across every source, not a collection per silo.
PII masking built in
Redact sensitive customer data at index time with one flag, before it reaches your models or logs.
Everything search needs,
tuned to your data.
Flagship
Captain Search
The fully-managed file search API.
Point it at your files and search them. Parsing, embeddings, vector DB and evaluation, all managed for you.
OCREmbedVector DBEvals
Main interface
Captain Evals
The main interface for Captain.
Call captain_eval from the hosted MCP server and it scores retrieval on evals written from your own files, then tunes toward the accuracy or latency target you set.
InspectEvalTuneReview
Managed sync
Captain Sync
Bucket change notifications drive the index in near real time, and scheduled reconciliation jobs sweep for anything a notification dropped.
Your buckets remain the source of truth.
S3R2GCSSupabaseBackblazeTerraform
Enterprise-grade
by default.
Security
TLS in transit, encryption at rest, and opt-in PII masking so sensitive data is hidden from your agents.
