| Web (paid APIs) |
|---|
| web_search | call | $0.040 | $0.030 | $0.020 | Standard web search via Exa with relevance-ranked results. It's the default tier and handles most agent lookups. |
| web_search_deep | call | $0.075 | $0.05625 | $0.0375 | Exa agentic deep search. It runs an iterative multi-step crawl for harder research questions a single lookup can't answer. |
| web_search_deep_reasoning | call | $0.100 | $0.075 | $0.050 | Exa deep search plus a reasoning pass that synthesizes across sources. Use this most thorough tier for complex research. |
| web_scrape | page | $0.010 | $0.0075 | $0.005 | Firecrawl-backed fetch that returns clean markdown for a single page. |
| Indexing |
|---|
| indexing_graphindex | 1k tokens | $0.010 | $0.0075 | $0.005 | GraphIndex builds an entity-relation graph over the corpus to power multi-hop queries. |
| indexing_doc2json | 1k tokens | $0.008 | $0.006 | $0.004 | Doc2JSON sliding-window extraction that turns long documents into structured records aligned to your schema. |
| indexing_pageindex | 1k tokens | $0.006 | $0.0045 | $0.003 | PageIndex builds a hierarchical tree over the document, which works best for reasoning over long structured docs. |
| metadata_enrichment | 1k tokens | $0.004 | $0.003 | $0.002 | Adds titles, summaries, and tags at indexing time so retrieval can filter and rerank on richer metadata. |
| indexing_chunkembed | 1k tokens | $0.002 | $0.0015 | $0.001 | ChunkEmbed does classic chunking plus embedding into pgvector. The default and cheapest strategy. |
| Extraction |
|---|
| ocr_pages | page | $0.006 | $0.0045 | $0.003 | VLM-based page extraction (LightOnOCR-2, Mistral, or PaddleOCR). We run it by default on every ingested document, so a single file with mixed scanned and text-extractable pages is handled correctly. If you know the source is all native text, opt into faster text-only parsers per ingestion run: OpenDataLoader, PyMuPDF (fitz), or pdfplumber. |
| Retrieval |
|---|
| tree_search | call | $0.010 | $0.0075 | $0.005 | PageIndex tree-walk retrieval, which navigates the document hierarchy instead of flat top-k chunks. |
| reranker_call | call | $0.010 | $0.0075 | $0.005 | Cross-encoder rerank via Cohere, Voyage, Jina, or ZeroEntropy, applied after retrieval to sharpen the top results. |
| hybrid_search | call | $0.004 | $0.003 | $0.002 | Vector + BM25 with reciprocal-rank fusion. This is our default retrieval mode. |
| query_enrichment | call | $0.004 | $0.003 | $0.002 | Query rewriting or HyDE, which expands the user's question before search to improve recall. |
| bm25_search | call | $0.002 | $0.0015 | $0.001 | BM25 / full-text search over your tables. Keyword-precise, and it needs no embeddings. |
| vector_search | call | $0.002 | $0.0015 | $0.001 | Plain cosine search over pgvector embeddings. It's the fastest path when you already know the question shape. |
| Agents & orchestrations |
|---|
| orchestration_run | run | $0.004 | $0.003 | $0.002 | Multi-agent orchestration base fee, covering the supervisor loop that routes between agents. |
| agent_run | run | $0.002 | $0.0015 | $0.001 | Base fee for a single ReAct agent invocation, before per-tool charges. |
| agent_tool_call | call | $0.002 | $0.0015 | $0.001 | One tool invocation from an agent, whether a custom HTTP tool, an MCP server, or a generic builtin. |
| agent_tool_code_execute | call | $0.002 | $0.0015 | $0.001 | Sandboxed code execution called from an agent. The runtime is isolated and there's no infra to manage. |
| Workflows |
|---|
| workflow_run | run | $0.002 | $0.0015 | $0.001 | Base fee per run of the visual / API workflow engine built into Powabase. |
| workflow_block_code | call | $0.002 | $0.0015 | $0.001 | A code block inside a workflow that runs your snippet in the sandboxed code executor. |
| workflow_block_external_api | call | $0.002 | $0.0015 | $0.001 | External HTTP block in a workflow, for outbound calls to third-party APIs. |