salvatore373/gaia_static_kb

GAIA Static Knowledge Base

Benchmark Reproducibility and Dynamic Web Search

Evaluating autonomous agents on benchmark datasets dependent on dynamic external services poses severe concerns on metric reliability. Specifically, non-stationary environments yield non-deterministic outputs across evaluation runs (i.e., they map the same input to possibly different outputs).

Consider an agent issuing a web search query at two distinct time points. At the earlier instant, the retrieved search results may require complex aggregation across multiple passages to answer the query. In the meantime, a new document directly answering the query may be indexed, enabling straightforward information extraction. The same query now retrieves a new web page, and the agent easily extracts the required information. Although the agent's intrinsic capabilities remain constant, the success rate on the benchmark dataset varies. Furthermore, transient network instability and server-side errors introduce additional variance across execution traces. Consequently, evaluation metrics relying on dynamic services may reflect lucky chances rather than true model performance.

To eliminate this source of variance, we decouple our evaluation on the GAIA benchmark from dynamic web search by constructing a deterministic, static knowledge base containing all the information required to solve the tasks.

Static Knowledge Base Construction

Corpus Aggregation

We first extracted execution traces from 32 (scaffold, model) configurations evaluated on GAIA, sourced from the HAL GAIA Leaderboard. We compute the set of unique URLs referenced across these traces and retrieve the corresponding web pages on a best-effort basis to form the GAIA static corpus. The resulting documents are stored in a relational database (located at gaia_static_kb/gaia_webpages_compr_ind_shards) with a single table web_pages and two columns: url, which is a unique index holding the web page's URL, and page_content storing the web page's content in plain text (compressed using zstd compression to save space on disk). These webpages make up the GAIA static corpus.

Vector Indexing

To support Retrieval-Augmented Generation, we index the GAIA static corpus into a dense vector database. This is saved at gaia_static_kb/gaia_vec_db. Documents are tokenized using the BAAI/bge-m3 tokenizer and segmented into contiguous chunks of at most 250 tokens to prevent retrieval dilution. Each chunk is enriched with metadata (document URL and title) and encoded as a dense vector using BAAI/bge-m3. Embeddings are indexed using FAISS, while an SQLite database maps each FAISS vector ID to its corresponding chunk text, title, and URL.