BrainHub/ rag

01 / Indexv1.0

Everything you know,
in one workspace,
cited and corrected.

Drop PDFs, images, audio and video into one workspace. Ask in plain language: a corrective RAG pipeline retrieves, grades, searches again when nothing fits, and shows the passages behind every answer.

PDFs·Images·Audio·Video

Open the workspace What makes it different

example answer

thread / 8a3f

How long do customers have to ask for a refund?

30 days under the 2026 pricing policy (p. 3). The 2025 policy allowed 14 days, so older receipts may follow the old rule.

Your sources disagree

Pricing-2026.pdf is current: it has the most recent document date.

Pricing-2026.pdfp.3Pricing-2025.pdfp.2

5 retrieved → 2 kept trail saved

Pipeline stages

Modalities

1024

Vector dimensions

Query variants per question

02 / Beyond citations

Cited is table stakes.
Here is what's next.

A citation says where a sentence came from. It does not say whether another document says the opposite, whether the answer would have been different last year, or what was left out. These three do, and the chat shows each one next to the answer it belongs to.

A / Disagreement

Sources that disagree, and which one is current

When the passages behind an answer come from two or more documents, a second model call checks whether they conflict. If they do, the answer carries the warning and names the documents; when their dates differ, the one with the most recent document date is marked current. That date comparison is plain code, not the model.

example, as the chat shows it

Your sources disagree

Refund window: 14 days in 2025, 30 days in 2026.

Pricing-2025.pdfPricing-2026.pdf · current

B / Time

Answer as of a date

Pick a date and only documents dated on or before it are searched: the date set at upload, or the upload day when none was set. Ask what the policy said in March 2025 and the 2026 revision stays out of the answer.

example, as the chat shows it

Answer as of2025-03-31

Pricing-2025.pdfdated 2025-01-15 · in
Pricing-2026.pdfdated 2026-02-01 · out

C / Trail

The decision trail

Every answer is saved with the path that produced it. “Why these answers?” shows the queries actually searched, each passage retrieved with its score and scale, which ones were kept, and whether the answer was flagged low confidence.

example, as the chat shows it

3 retrieved → 2 keptreranked

Pricing-2026.pdf p.30.91 · kept
Pricing-2025.pdf p.20.84 · kept
Handbook.pdf p.70.31 · dropped

03 / Pipeline

Five stages.
No shortcuts.

Every question follows the same path. When nothing clears the relevance bar, the question is rewritten and searched once more. If that still fails, the model is told the passages may not answer it, and the answer is marked low confidence.

0101
Multi-modal ingest
PDF text is read with pypdf; pages with no text layer are rendered as images and embedded as such. Images and video are embedded directly by Voyage, audio is transcribed by Deepgram. libmagic sniffs MIME from the bytes, not the extension.
libmagic · pypdf + PyMuPDF · max 20MB
0202
Embed & index
One multimodal Voyage model embeds documents and queries, so text, images and video share a single vector space. A Postgres tsvector index is kept alongside for keyword search.
voyage-multimodal-3.5 · 1024d · pgvector HNSW
0303
Hybrid retrieval
Each question becomes a few search queries. For each one, a pgvector search and a Postgres full-text search are fused with Reciprocal Rank Fusion; a Cohere reranker then orders the pool and keeps the top five.
rrf_k=60 · top_k=5 · rerank=5
0404
Relevance gate
The Cohere reranker scores each passage from 0 to 1; passages at 0.7 or above go to the answer. If none clears the bar, the question is rewritten and the library searched once more. If it still fails, the two best passages go through anyway and the answer is marked low confidence. Without the reranker, the fused top results pass as they are, and fewer than two counts as low confidence.
cohere ≥ 0.7 · rewrite + search again
0505
Grounded synthesis
Claude Sonnet is instructed to answer only from the retrieved passages, cite document and page, and say when they do not contain the answer instead of filling the gap from general knowledge. Tokens stream as they are generated.
claude-sonnet-5 · stream=true

04 / Engineering

Built where the demos stop.

six choices that matter

01 / Isolation

Tenant-isolated retrieval

Every chunk row carries a user_id; vector and keyword queries WHERE-filter on it before fusion or rerank. No cross-tenant path exists at the SQL layer.

02 / LLM

Multi-provider routing

Anthropic native or any OpenRouter model, chosen by env var. Query generation, the disagreement check and the answer all go through that one client; embeddings and reranking stay on Voyage and Cohere.

03 / Fallback

Web search is opt-in

Sending a question to a third party is off by default. When enabled, a Tavily search runs only after the rewritten query still finds nothing, never with a date cut-off, and web results are marked as web wherever they appear.

04 / Citations

Answers show their passages

Each answer lists the passages it was generated from, with document, page, date and a snippet; a click opens the file at that page. The prompt instructs the model to answer only from those passages and to say so when they do not contain the answer.

05 / Workspace

Private workspace per account

Documents, retrieval, chat threads and decision trails are scoped to the account that created them.

06 / Upload

libmagic byte sniffing

MIME is read from bytes, not file extension. Allowlist enforced server-side; renamed payloads are rejected before they touch storage, and the size limit is counted while the upload streams in.

05 / Stackno fluff

Claude Sonnet●Voyage multimodal-3.5●Cohere Rerank●pgvector HNSW●PostgreSQL●Redis 7●arq●FastAPI●Next.js 16●React 19●Turbopack●OpenAI Realtime●Deepgram Nova-3●PyMuPDF●libmagic●

Get started

Stop grepping
your own files.

Drop what you have into one workspace and ask. Every answer shows the passages it came from, and the path it took to find them.

$brainhubinit--workspacemy-workspace

Don't have an account?

BrainHub/ rag

01 / Indexv1.0

Everything you know,
in one workspace,
cited and corrected.

PDFs·Images·Audio·Video

Open the workspace What makes it different

example answer

thread / 8a3f

How long do customers have to ask for a refund?

30 days under the 2026 pricing policy (p. 3). The 2025 policy allowed 14 days, so older receipts may follow the old rule.

Your sources disagree

Pricing-2026.pdf is current: it has the most recent document date.

Pricing-2026.pdfp.3Pricing-2025.pdfp.2

5 retrieved → 2 kept trail saved

Pipeline stages

Modalities

1024

Vector dimensions

Query variants per question

02 / Beyond citations

Cited is table stakes.
Here is what's next.

A / Disagreement

Sources that disagree, and which one is current

example, as the chat shows it

Your sources disagree

Refund window: 14 days in 2025, 30 days in 2026.

Pricing-2025.pdfPricing-2026.pdf · current

B / Time

Answer as of a date

example, as the chat shows it

Answer as of2025-03-31

Pricing-2025.pdfdated 2025-01-15 · in
Pricing-2026.pdfdated 2026-02-01 · out

C / Trail

The decision trail

example, as the chat shows it

3 retrieved → 2 keptreranked

Pricing-2026.pdf p.30.91 · kept
Pricing-2025.pdf p.20.84 · kept
Handbook.pdf p.70.31 · dropped

03 / Pipeline

Five stages.
No shortcuts.

0101
Multi-modal ingest
PDF text is read with pypdf; pages with no text layer are rendered as images and embedded as such. Images and video are embedded directly by Voyage, audio is transcribed by Deepgram. libmagic sniffs MIME from the bytes, not the extension.
libmagic · pypdf + PyMuPDF · max 20MB
0202
Embed & index
One multimodal Voyage model embeds documents and queries, so text, images and video share a single vector space. A Postgres tsvector index is kept alongside for keyword search.
voyage-multimodal-3.5 · 1024d · pgvector HNSW
0303
Hybrid retrieval
Each question becomes a few search queries. For each one, a pgvector search and a Postgres full-text search are fused with Reciprocal Rank Fusion; a Cohere reranker then orders the pool and keeps the top five.
rrf_k=60 · top_k=5 · rerank=5
0404
Relevance gate
The Cohere reranker scores each passage from 0 to 1; passages at 0.7 or above go to the answer. If none clears the bar, the question is rewritten and the library searched once more. If it still fails, the two best passages go through anyway and the answer is marked low confidence. Without the reranker, the fused top results pass as they are, and fewer than two counts as low confidence.
cohere ≥ 0.7 · rewrite + search again
0505
Grounded synthesis
Claude Sonnet is instructed to answer only from the retrieved passages, cite document and page, and say when they do not contain the answer instead of filling the gap from general knowledge. Tokens stream as they are generated.
claude-sonnet-5 · stream=true

04 / Engineering

Built where the demos stop.

six choices that matter

01 / Isolation

Tenant-isolated retrieval

Every chunk row carries a user_id; vector and keyword queries WHERE-filter on it before fusion or rerank. No cross-tenant path exists at the SQL layer.

02 / LLM

Multi-provider routing

03 / Fallback

Web search is opt-in

04 / Citations

Answers show their passages

05 / Workspace

Private workspace per account

Documents, retrieval, chat threads and decision trails are scoped to the account that created them.

06 / Upload

libmagic byte sniffing

MIME is read from bytes, not file extension. Allowlist enforced server-side; renamed payloads are rejected before they touch storage, and the size limit is counted while the upload streams in.

05 / Stackno fluff

Get started

Stop grepping
your own files.

Drop what you have into one workspace and ask. Every answer shows the passages it came from, and the path it took to find them.

$brainhubinit--workspacemy-workspace

Don't have an account?

Everything you know,in one workspace,cited and corrected.

Cited is table stakes.Here is what's next.

Sources that disagree, and which one is current

Answer as of a date

The decision trail

Five stages.No shortcuts.

Multi-modal ingest

Embed & index

Hybrid retrieval

Relevance gate

Grounded synthesis

Built where the demos stop.

Tenant-isolated retrieval

Multi-provider routing

Web search is opt-in

Answers show their passages

Private workspace per account

libmagic byte sniffing

Stop greppingyour own files.

Everything you know,in one workspace,cited and corrected.

Cited is table stakes.Here is what's next.

Sources that disagree, and which one is current

Answer as of a date

The decision trail

Five stages.No shortcuts.

Multi-modal ingest

Embed & index

Hybrid retrieval

Relevance gate

Grounded synthesis

Built where the demos stop.

Tenant-isolated retrieval

Multi-provider routing

Web search is opt-in

Answers show their passages

Private workspace per account

libmagic byte sniffing

Stop greppingyour own files.

Everything you know,
in one workspace,
cited and corrected.

Cited is table stakes.
Here is what's next.

Five stages.
No shortcuts.

Stop grepping
your own files.

Everything you know,
in one workspace,
cited and corrected.

Cited is table stakes.
Here is what's next.

Five stages.
No shortcuts.

Stop grepping
your own files.