M5.1.6 arxiv-rag: cost ledger fields (`embedding_calls`) #43

New issue

Open

opened 2026-04-08 23:17:59 +00:00 by claude-code · 0 comments

claude-code commented

2026-04-08 23:17:59 +00:00

Collaborator

Final sub-milestone of Issue #37. Design: ArxivRagProposal.

Goal

Extend the cost ledger so arxiv-rag calls record their embedding usage alongside the existing token / Tavily fields. Closes the per-researcher cost-tracking loop.

Scope

obs/costs.py:
- Add embedding_calls: int = 0 and embedding_model: Optional[str] = None to ledger entries
- Extend PriceTable.estimate_call_usd to accept an embedding component:
  - For local embedding models, contribute $0
  - For API embedding models (e.g. voyage-3), look up [embeddings."<model>"] in prices.toml and bill per-token
- Seed prices.toml with [embeddings."nomic-embed-text-v1.5"] (free) and [embeddings."voyage-3"] (placeholder rate)
researchers/arxiv/agent.py records the number of retrieve_chunks calls and embedding model name, passes them through when calling ledger.record(...)
marchwarden costs rendering: include embedding_calls in the per-day / per-model breakdowns where it's nonzero

Tests

Local embedding model: ledger entry has embedding_calls > 0, estimated_cost_usd only counts the synthesis tokens
API embedding model with a price-table entry: estimated cost includes the embedding component
Existing web researcher entries still work (no embedding fields → costs render the same as today)

Branch

feat/arxiv-rag-cost-ledger

Blocked by: M5.1.5. Closes the M5.1 milestone.

Final sub-milestone of Issue #37. Design: [ArxivRagProposal](https://forgejo.labbity.unbiasedgeek.com/archeious/marchwarden/wiki/ArxivRagProposal). ## Goal Extend the cost ledger so arxiv-rag calls record their embedding usage alongside the existing token / Tavily fields. Closes the per-researcher cost-tracking loop. ## Scope - `obs/costs.py`: - Add `embedding_calls: int = 0` and `embedding_model: Optional[str] = None` to ledger entries - Extend `PriceTable.estimate_call_usd` to accept an embedding component: - For local embedding models, contribute $0 - For API embedding models (e.g. `voyage-3`), look up `[embeddings."<model>"]` in `prices.toml` and bill per-token - Seed prices.toml with `[embeddings."nomic-embed-text-v1.5"]` (free) and `[embeddings."voyage-3"]` (placeholder rate) - `researchers/arxiv/agent.py` records the number of `retrieve_chunks` calls and embedding model name, passes them through when calling `ledger.record(...)` - `marchwarden costs` rendering: include embedding_calls in the per-day / per-model breakdowns where it's nonzero ## Tests - Local embedding model: ledger entry has `embedding_calls > 0`, `estimated_cost_usd` only counts the synthesis tokens - API embedding model with a price-table entry: estimated cost includes the embedding component - Existing web researcher entries still work (no embedding fields → costs render the same as today) ## Branch `feat/arxiv-rag-cost-ledger` Blocked by: M5.1.5. Closes the M5.1 milestone.

claude-code changed title from ~~A.6 arxiv-rag: cost ledger fields (embedding_calls)~~ to M5.1.6 arxiv-rag: cost ledger fields (embedding_calls)

2026-04-08 23:23:15 +00:00