BatchHive AI Studio V12.31.12 — Relationship Rows

BatchHive AI Studio · V12.31.12

Operate many models.Keep every row under control.

BatchHive combines two focused AI surfaces in one browser runtime: Studio for high-volume, row-aligned batch processing over Excel and CSV datasets, and Chat for interactive multimodal work. Both share the same provider catalog, model pool, and theme system without forcing batch execution and conversation into the same interface.

Studio + Chat surfaces
Shared model pool
Row-aligned batch outputs
Checkpoints & provenance

Platform

One shared foundation. Two focused product surfaces.

Studio and Chat use the same provider catalog, model pool, network layer, search integrations, key storage, and theme system. Their interfaces stay separate so batch operations and conversational work can evolve without competing for the same screen.

Multi-model operations

Bring hosted APIs, local endpoints, and Puter into one model pool. Run a single model, compare many side by side, or route work dynamically without rebuilding the dataset.

Model poolCompare / Multi-RunLocal endpoints

Dataset-native execution

Work directly from spreadsheet rows instead of copy/paste prompts. Variable mapping, row identity, large-dataset storage, and incremental results stay attached to the source data.

Excel / CSVRow identityIndexedDB

Capacity-aware Smart Route

Admission happens before a row is claimed. Routing reacts to RPM, TPM, Retry-After, provider cooldowns, daily limits, and available capacity across the selected lanes.

PreflightFailoverAdaptive capacity

Traceable results

Outputs remain associated with their source rows and execution context. Provenance, route attempts, failure diagnostics, and resumable checkpoints make long runs inspectable.

ProvenanceCheckpointsFailure diagnostics

Product areas

Choose the interface that matches the job.

Studio is optimized for deterministic, row-based operations. Chat is optimized for conversation, multimodal context, search, and iterative reasoning. Moving between them never requires a reload.

01Batch operations

Studio

Configure providers, load spreadsheet data, map variables, define the prompt once, tune execution controls, run at scale with Smart Route or Compare mode, and inspect row-aligned results with provenance, checkpoints, and exports.

6-stage workflowSmart RouteCompare modeLarge Dataset modeResults & provenance

02Conversation & tools

Chat

Work interactively with the shared model pool: persistent and temporary conversations, branches and message versions, multimodal attachments, reasoning and context controls, manually selected web search, global memory, and Markdown/JSON import & export.

Conversation historyAttachmentsWeb SearchMemoryShared model pool

Studio workflow

Six stages. One continuous batch context.

Jump directly to any stage inside Studio. Provider configuration, dataset mapping, prompt controls, routing state, and results remain connected without turning the interface into one oversized form.

Providers & integrations

One model pool across hosted, free, and local providers.

Bring your own API keys, sign in to Puter for free hosted access, or point BatchHive at your local endpoints. Smart Route and Compare mode treat every selected lane as one capacity pool, and provider calls go straight from your browser to the provider.

Hosted model APIs · bring your own key

Google GeminiOfficial catalog + account models
OpenRouterUnified gateway · BYOK model discovery
GroqHigh-throughput inference
Mistral AIEuropean model family
CerebrasFast inference lanes
SambaNovaHosted inference
NVIDIA NIMHosted inference
DeepSeekReasoning models
AnthropicClaude model family
OpenAIGPT model family
Dahl InferenceCommunity gateway

Free & gateway lanes

PuterFree access to 500+ hosted models with a Puter account — no API keys. Isolated pool with its own results and checkpoints.
NaraRouterOpenAI-compatible gateway with free-plan aliases and optional authenticated refresh.
PollinationsFree community lane with per-route pricing awareness.

Local endpoints

OllamaModels on your machine — local capability discovery at localhost:11434.
LM StudioLocal server models with live catalog discovery at localhost:1234.

Web search tooling · priority-ordered fallback

OpenRouter Web SearchProvider-native search grounding
FirecrawlCrawl-grounded results with your own key
NeuroAPISearch fallback lane
Source verificationOptional verification pass on retrieved sources

Scale & reliability

Designed for work that outlives a single request.

Large datasets introduce memory, quota, retry, and traceability problems. BatchHive treats those as execution concerns instead of leaving them to the prompt.

Bounded browser memory. Disk-backed progress.

Large Dataset mode parses and stores source rows in IndexedDB, executes bounded windows, persists incremental results, and keeps checkpoints resumable. The active window stays small even when the source file is not.

IndexedDBdisk-backed rows & results

Bounded windowscontrolled browser memory

Streaming exportlong-run friendly output

Admission before row claim

Capacity checks happen before work is assigned to an execution lane.

Retry & reroute isolation

Transient failures can release rows to another eligible provider or model.

Stable row identity

Transport IDs and persisted ordinals keep outputs bound to the intended source row.

Security & data handling

Browser-first by design. Visible instead of hidden.

BatchHive has no application backend: provider, dataset, chat, and search traffic runs from your browser directly to the services you configure. Datasets, results, checkpoints, and conversations are stored in your browser's storage — including disk-backed IndexedDB for large datasets — not on a BatchHive server.

Provider keys are convenience storage, not a secrets vault: saved keys are obfuscated in local or session storage, never presented as encrypted, and can be cleared in one action. Restricted keys and trusted devices remain the right operating practice. When the app is hosted on Puter, network requests may fall back to Puter's networking relay, and that behavior stays visible in transport diagnostics.

Direct provider calls

Requests go from your browser to the configured provider; there is no BatchHive middleman server that sees your prompts or data.

Keys stay on the device

Session-only by default, optional persistence, XOR obfuscation in storage, and a one-click "forget everything" control.

Data you can take with you

Results export to Excel/CSV, conversations export to Markdown/JSON, and large runs stream their output incrementally.

Inspectable execution

Route attempts, retry budgets, cooldowns, quota scopes, and provenance stay visible in the Run and Results workspaces.

Choose your surface

Run production batches in Studio. Think interactively in Chat.

Both surfaces share the same BatchHive foundation, so you can move between structured execution and conversation without rebuilding provider or model configuration.

API Provider

Official model catalog

Text / chat only Built-in fallback

Choose a provider to load its official model catalog.

Select a model to see context, capabilities, pricing/catalog metadata, and source.

Provider Key

INPUT

Dataset

Upload and inspect the source file, then move on. This remains a focused setup step—not the main workspace.

▦ Dataset Source

Upload Excel or CSV as a structured dataset The workbook stays structured. Rows are never flattened into the text box.

No file loaded 0 rows 0 columns 0 ready

Sheet / Table

Header Row

No dataset loaded. Preview shows up to 15 records

Upload a dataset to preview rows.

Advanced quick-test mode. Each non-empty line is one record. Every prompt variable resolves to that line, so no column mapping is required.

Manual Records

Manual mode is for quick tests. For production files, use File Dataset.

◈ Data Analysis

✎ Prompt & Variable Mapping

Prompt Template

Empty Variable Cells

Choose a template or write your own prompt.

Prompt Template Use {{variable}} placeholders

Insert a column variable:

Variable → Column MappingNo variables detected.

Add {{variables}} to the prompt. In Manual mode, variables resolve to each text line automatically.

↑ Output Columns

Result Mode

Output Column Header

Text Export Format

◈ Prompt Preview Preview first records after variable substitution

Upload a dataset and map variables to preview the final per-row prompt.

▣ Run PlanSingle model · 1 run

Add models from Provider & Models

Run Plan

One model stays Single automatically; adding more models or runs switches to Compare.

Choose a model to prepare a single run.

Selected Model

Choose a provider/model above, then add it to the comparison plan.

Runs per add

Shared Model Pool is empty.

⚡ Smart Route

Every source row is processed once. Routes share the queue; each route can follow Quick Controls or use its own custom values.

＋

Add routes from Provider & ModelsPress Add on any model row. The same model is added to Compare and Smart Route together.

Routing Policy

Default Weight

Re-route failed batchesOn a failed attempt, unresolved rows immediately move to a different eligible model/provider.

Route Pool

No routes added yet.

Active routes0

Workers0

Fastest Available
Failover enabled

ADVANCED

Processing Settings

Token budget and optional batch protection.

◈ Output Token Budget

Max Output Tokens

Effective Budget

Auto estimate will appear here.

Auto mode estimates the completion budget from task/output shape and batch size. Large static classifier taxonomies do not inflate the completion budget. Token-limit retries can temporarily raise the request budget.

🛡 Adaptive Batch Protection: OFF

Optional. Batch Size stays exactly as configured until you enable the context-window guard.

Optional

DEFAULT CONTROLS PROMPT CACHE ON

Batch Size / Run

Concurrency / Run

Rate Limit Retry Wait

Retries (Single / Compare)

Web Search

Search Provider

Fallback Priority

→

Verify Sources ONValidate final relationship sources

Priority: Firecrawl → NeuroAPI → OpenRouter · External providers attach ranked evidence to any LLM route; missing/exhausted external keys are skipped. OpenRouter Web Search uses its web plugin and requires an OpenRouter route.

Firecrawl API Key

Available in search priority

TRUE BATCH · prompt once/request

Temperature

Circuit Breaker — max consecutive errors

Rate-limit policy: Compare/Single waits then retries the same run; Smart Route releases the current row immediately, cools only the throttled route, then returns that route to the pool automatically. Auto uses provider Retry-After when available, otherwise 60s.

▶ Run

Total

Done ✓

Errors

Retried

—

Rows/min

—

ETA

Ready

◈ Output Quality Score

— Run a batch to see metrics

◈ Live Log

Current activity

Circuit OK

Ready idle

↑ Output

Results stay aligned to source rows. Failed compare cells remain blank.

Run a job to preview results.

Raw text output

— output will appear here —

Unavailable

Context0 / —

◈

Start a new conversation

Choose a model from the shared BatchHive Model Pool, then send a message.

Chat is independent from datasets, prompts, runs, and results.