CSR
Citation Survival Rate, a measure of whether citations to correct sources survive across the MCP boundary, combining support checks and source-tracing checks.
Timeline (last 90 days)
CSR (compressed sparse row) is a matrix format that stores only non-zero elements and their column indices, represented by three arrays—values, column indices, and row offsets—with space overhead proportional to the number of edges.
GitBlast uses an AST lexical tokenizer in the browser to parse code structure into a low-level compressed sparse row (CSR) in-memory matrix.
AI crawlers process pages within strict time budgets and are unfriendly to client-side-rendered (CSR) single-page applications; with a pure CSR architecture and no SSR or prerendering, crawlers may only get an empty mount node.
On 127 FastAPI documentation evaluation questions (82 answerable from a single page, 29 spanning pages, 16 not covered by the documentation), the Agent built with the OpenAI Agents SDK under the default API had a first-round CSR of only 0.323, while the LangGraph Agent was 0.614 and single-pass RAG was 0.618.
Of the 127 first-round OpenAI Agents SDK answers, 66 did not use the required [url] format, but reading them one by one found that 51 cited sources as Markdown links and 7 used other bracket formats, with only 8 giving no URL at all; after correcting the scoring definition, the SDK Agent's CSR rose from 0.323 to 0.470.
Three repeated experiments show that about sixty percent of the CSR gap between the OpenAI Agents SDK and single-pass RAG comes from citation format mismatch and forty percent is genuine citation loss; the gap was −0.12 under the original scorer, −0.09 under gpt-4o judging, and about −0.30 when only the specified format is accepted.
In the evaluation, gpt-4o judging had the highest agreement with the author's own 115 blind judgments (κ 0.86), and CSR judgments matched human judgment on 25 to 27 of 30 blind cases.
Running the experiment only once is not enough; single-run results for the remaining gap may fall between −0.05 and −0.16, and multiple repeated experiments are needed to assess citation loss stably.
Tagging each retrieved block with a short source ID (such as [S3fa21c]), requiring the Agent to cite with [ID], and mapping back to URLs after generation reduced format-error answers from about 50 per round to about 5 on the self-hosted mirror, raised the SDK Agent's CSR by 0.10 to 0.13, and eliminated the gap with single-pass RAG.
CSR (Citation Survival Rate) requires each citation to pass two checks: a support check that uses FActScore-style atomic claim decomposition plus an ALCE-style entailment model to verify whether the cited block supports the claim, and a provenance check that relies on tool-call logs to confirm the URL actually appeared in the content the Agent retrieved.
2 more timeline events
All Facts (12)
CSR (compressed sparse row) is a matrix format that stores only non-zero elements and their column indices, represented by three arrays—values, column indices, and row offsets—with space overhead proportional to the number of edges.
50%UnverifiedGitBlast uses an AST lexical tokenizer in the browser to parse code structure into a low-level compressed sparse row (CSR) in-memory matrix.
50%UnverifiedAI crawlers process pages within strict time budgets and are unfriendly to client-side-rendered (CSR) single-page applications; with a pure CSR architecture and no SSR or prerendering, crawlers may only get an empty mount node.
50%UnverifiedOn 127 FastAPI documentation evaluation questions (82 answerable from a single page, 29 spanning pages, 16 not covered by the documentation), the Agent built with the OpenAI Agents SDK under the default API had a first-round CSR of only 0.323, while the LangGraph Agent was 0.614 and single-pass RAG was 0.618.
50%UnverifiedOf the 127 first-round OpenAI Agents SDK answers, 66 did not use the required [url] format, but reading them one by one found that 51 cited sources as Markdown links and 7 used other bracket formats, with only 8 giving no URL at all; after correcting the scoring definition, the SDK Agent's CSR rose from 0.323 to 0.470.
50%UnverifiedThree repeated experiments show that about sixty percent of the CSR gap between the OpenAI Agents SDK and single-pass RAG comes from citation format mismatch and forty percent is genuine citation loss; the gap was −0.12 under the original scorer, −0.09 under gpt-4o judging, and about −0.30 when only the specified format is accepted.
50%UnverifiedIn the evaluation, gpt-4o judging had the highest agreement with the author's own 115 blind judgments (κ 0.86), and CSR judgments matched human judgment on 25 to 27 of 30 blind cases.
50%UnverifiedRunning the experiment only once is not enough; single-run results for the remaining gap may fall between −0.05 and −0.16, and multiple repeated experiments are needed to assess citation loss stably.
50%UnverifiedTagging each retrieved block with a short source ID (such as [S3fa21c]), requiring the Agent to cite with [ID], and mapping back to URLs after generation reduced format-error answers from about 50 per round to about 5 on the self-hosted mirror, raised the SDK Agent's CSR by 0.10 to 0.13, and eliminated the gap with single-pass RAG.
50%UnverifiedCSR (Citation Survival Rate) requires each citation to pass two checks: a support check that uses FActScore-style atomic claim decomposition plus an ALCE-style entailment model to verify whether the cited block supports the claim, and a provenance check that relies on tool-call logs to confirm the URL actually appeared in the content the Agent retrieved.
50%UnverifiedThe CSR denominator counts only the correct sources the Agent actually retrieved; retrieval failures are not counted in CSR, and CSR measures only what happens to correct sources after they cross the MCP boundary.
50%UnverifiedThe query_start_loc array uses prefix-sum encoding and is a variant of the Compressed Sparse Row (CSR) format, used to allow the attention kernel to distinguish boundaries between different requests.
50%