You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
- vfb_list_connectome_datasets / vfb_query_connectivity: comparative class-level or neuron-level connectivity across datasets
1475
1587
- search_reviewed_docs: search approved VFB, NeuroFly, VFB Connect docs, and reviewed FlyBase pages using a server-side site index
1476
1588
- get_reviewed_page: fetch and extract content from an approved page returned by search_reviewed_docs
1477
1589
- search_pubmed / get_pubmed_article: search and fetch peer-reviewed publications
@@ -1480,15 +1592,26 @@ TOOLS:
1480
1592
${VFB_QUERY_LINK_SKILL}
1481
1593
1482
1594
TOOL SELECTION:
1595
+
- Choose tools dynamically based on the user request and available evidence; the guidance below is preferred, not a rigid workflow.
1483
1596
- Questions about VFB terms, anatomy, neurons, genes, or datasets: use VFB tools
1597
+
- For VFB entity questions where suitable query types are available, prefer vfb_get_term_info + vfb_run_query as a first pass because vfb_run_query is usually cached and faster.
1598
+
- Questions about FlyBase genes/alleles/insertions/stocks: use vfb_resolve_entity first (if unresolved), then vfb_find_stocks
1599
+
- Questions about split-GAL4 combination names/synonyms (for example MB002B, SS04495): use vfb_resolve_combination first, then vfb_find_combo_publications (and optionally vfb_find_stocks if the user asks for lines)
1600
+
- Questions about comparative connectivity between neuron classes across datasets: use vfb_query_connectivity (optionally vfb_list_connectome_datasets first to pick valid dataset symbols)
1484
1601
- Questions about published papers or recent literature: use PubMed first, optionally bioRxiv/medRxiv for preprints
1485
1602
- Questions about VFB, NeuroFly, VFB Connect Python documentation, or approved FlyBase documentation pages, news posts, workshops, conference pages, or event dates: use search_reviewed_docs, then use get_reviewed_page when you need page details
1486
1603
- For questions about how to run VFB queries in Python or how to use vfb-connect, prioritize search_reviewed_docs/get_reviewed_page on vfb-connect.readthedocs.io alongside VFB tool outputs when useful.
1487
-
- For connectivity, synaptic, or NBLAST questions, and especially when the user explicitly asks for vfb_run_query, do not use reviewed-doc search first; use VFB tools (vfb_search_terms/vfb_get_term_info/vfb_run_query).
1604
+
- For connectivity, synaptic, or NBLAST questions, and especially when the user explicitly asks for vfb_run_query, do not use reviewed-doc search first; use VFB tools (vfb_search_terms/vfb_get_term_info/vfb_run_query). Use vfb_query_connectivity when the user asks for class-to-class connectivity comparisons across datasets.
1488
1605
- Do not attempt general web search or browsing outside the approved reviewed-doc index
1489
1606
1607
+
ENTITY RESOLUTION RULES:
1608
+
- If vfb_resolve_entity or vfb_resolve_combination returns match_type SYNONYM or BROAD, confirm the resolved entity with the user before running downstream tools.
1609
+
- If resolver output includes multiple candidates, show a short disambiguation list and ask the user to choose before continuing.
1610
+
- If the user already provided a canonical FlyBase ID (for example FBgn..., FBal..., FBti..., FBco..., FBst...), you may call downstream tools directly.
1611
+
1490
1612
TOOL ECONOMY:
1491
1613
- Prefer the fewest tool steps needed to produce a useful answer.
1614
+
- Start with cached vfb_run_query pathways when they can answer the question, then use other tools for deeper refinement only when needed.
1492
1615
- Do not keep calling tools just to exhaustively enumerate large result sets.
1493
1616
- If the question is broad or combinatorial, stop once you have enough evidence to give a partial answer.
1494
1617
- For broad gene-expression or transgene-pattern requests, prefer a short representative list (about 3-5 items) and ask how the user wants to narrow further instead of trying to enumerate everything in one turn.
@@ -1511,9 +1634,13 @@ TOOL RELAY:
1511
1634
- If a question needs data and no results are available yet, request tools first, then answer after results arrive.
1512
1635
1513
1636
FOLLOW-UP QUESTIONS:
1514
-
When useful, suggest 2-3 short follow-up questions relevant to Drosophila neuroscience and actionable in this chat.`
1637
+
When useful, suggest 2-3 short potential follow-up questions that are directly answerable with the available tools in this chat.`
'- For this request, prioritize VFB tools over reviewed-doc search.',
1704
-
'- Use vfb_search_terms and/or vfb_get_term_info to identify the target entity and valid query types.',
1705
-
'- Use vfb_run_query when a relevant query_type is available.'
1825
+
'- Choose the smallest set of tools that best answers the user request.',
1826
+
'- For VFB query-type questions, prefer vfb_get_term_info + vfb_run_query as the first pass because vfb_run_query is typically cached and fast.',
1827
+
'- Use more specialized tools (for example vfb_query_connectivity, vfb_resolve_entity, vfb_find_stocks, vfb_resolve_combination, vfb_find_combo_publications) when deeper refinement is needed.',
1828
+
'- Prefer direct data tools over documentation search when the question asks for concrete VFB data.',
1829
+
'- If existing tool outputs already answer the question, provide the final answer instead of requesting more tools.'
1706
1830
]
1707
1831
1708
1832
if(explicitRunQueryRequested){
1709
1833
policyBullets.push('- The user explicitly asked for vfb_run_query, so include a plan that leads to vfb_run_query.')
1710
1834
}
1711
1835
1712
1836
if(connectivityIntent){
1713
-
policyBullets.push('- This is a connectivity-style request; do not default to search_reviewed_docs first.')
1837
+
policyBullets.push('- This is a connectivity-style request; favor VFB connectivity/query tools over docs-only search.')
1714
1838
}
1715
1839
1716
1840
if(missingRunQueryExecution){
@@ -1974,6 +2098,7 @@ Using only the gathered tool outputs already provided in this conversation:
1974
2098
- clearly say that the answer is partial because the request branched into too many tool steps
1975
2099
- summarize the strongest findings you already have
1976
2100
- end with 2-4 direct clarification questions the user can answer so you can continue in a narrower, lower-tool way
2101
+
- make those questions concrete and answerable with the tools available in this chat
1977
2102
1978
2103
Do not call tools. Do not ask to browse the web.`
1979
2104
@@ -2016,6 +2141,7 @@ Using only the existing conversation and any tool outputs already provided:
2016
2141
- give a brief summary of what direction is available so far
2017
2142
- do not invent missing facts
2018
2143
- ask 2-4 short clarifying questions the user can answer so the next turn can be narrower and easier to resolve
2144
+
- keep clarifying questions concrete and answerable with the tools available in this chat
2019
2145
2020
2146
Do not call tools. Do not ask to browse the web.`
2021
2147
@@ -2062,6 +2188,7 @@ Using only the existing conversation, any tool outputs already provided, and any
2062
2188
- if the evidence is still too incomplete, say that briefly and ask 2-4 short clarifying questions
2063
2189
- prefer a short concrete answer over more questions if the available evidence already supports one
2064
2190
- do not invent missing facts
2191
+
- if you ask questions, make them concrete and answerable with the tools available in this chat
2065
2192
2066
2193
Do not call tools. Do not ask to browse the web.`
2067
2194
@@ -2164,10 +2291,8 @@ async function processResponseStream({
0 commit comments