') + ')', 'gi'); if (regex.test(text)) { found = true; var frag = document.createDocumentFragment(); var parts = text.split(regex); parts.forEach(function(part, i) { if (i % 2 === 0) { frag.appendChild(document.createTextNode(part)); } else { var span = document.createElement('span'); span.className = 'userscript-highlight'; span.textContent = part; frag.appendChild(span); } }); node.parentNode.replaceChild(frag, node); } }); } else if (node.nodeType === 1 && node.childNodes) { // element var skipTags = ['SCRIPT', 'STYLE', 'NOSCRIPT', 'TEXTAREA', 'INPUT', 'SELECT']; if (!skipTags.includes(node.tagName)) { Array.from(node.childNodes).forEach(highlight); } } } highlight(document.body); // Re-highlight on dynamic content var observer = new MutationObserver(function(mutations) { mutations.forEach(function(m) { m.addedNodes.forEach(function(node) { if (node.nodeType === 1 || node.nodeType === 3) highlight(node); }); }); }); observer.observe(document.body, { childList: true, subtree: true }); })(); } } catch(__e) { console.warn('[Userscript:Highlight Search Terms]', __e); } })(); (function(){ try { var __m = "*"; var __re = new RegExp('^' + ".*" + ', 'i'); if (__m === '*' || __re.test(location.href)) { // Strip utm_, fbclid, gclid, etc. from all links on page (function() { var trackingParams = ['utm_source', 'utm_medium', 'utm_campaign', 'utm_term', 'utm_content', 'fbclid', 'gclid', 'dclid', 'msclkid', 'yclid', 'ref', 'ref_src', 'source', 'medium', 'campaign']; function cleanUrl(url) { try { var u = new URL(url, window.location.origin); var changed = false; trackingParams.forEach(function(p) { if (u.searchParams.has(p)) { u.searchParams.delete(p); changed = true; } }); return changed ? u.toString() : url; } catch (e) { return url; } } function cleanLinks() { document.querySelectorAll('a[href]').forEach(function(a) { var clean = cleanUrl(a.href); if (clean !== a.href) a.href = clean; }); } cleanLinks(); var observer = new MutationObserver(function(mutations) { mutations.forEach(function(m) { m.addedNodes.forEach(function(node) { if (node.nodeType === 1) { if (node.tagName === 'A') cleanLinks(); node.querySelectorAll('a[href]').forEach(function(a) { var clean = cleanUrl(a.href); if (clean !== a.href) a.href = clean; }); } }); }); }); observer.observe(document.body, { childList: true, subtree: true }); })(); } } catch(__e) { console.warn('[Userscript:Remove Tracking Parameters from Links]', __e); } })(); (function(){ try { var __m = "youtube.com"; var __re = new RegExp('^' + "youtube\\.com" + ', 'i'); if (__m === '*' || __re.test(location.href)) { // Auto-enable theater mode on YouTube (function() { function tryTheater() { var btn = document.querySelector('button[aria-label="Theater mode"], ytd-player #player button[title="Theater mode"]'); if (btn && !btn.classList.contains('activated')) { btn.click(); } } // Try immediately tryTheater(); // Try after navigation (SPA) var lastUrl = location.href; setInterval(function() { if (location.href !== lastUrl) { lastUrl = location.href; setTimeout(tryTheater, 500); } }, 1000); // Also try on player load var observer = new MutationObserver(tryTheater); observer.observe(document.body, { childList: true, subtree: true }); })(); } } catch(__e) { console.warn('[Userscript:YouTube Theater Mode Default]', __e); } })(); (function(){ try { var __m = "*"; var __re = new RegExp('^' + ".*" + ', 'i'); if (__m === '*' || __re.test(location.href)) { // Remove or un-stick sticky/fixed headers that block content (function() { function unstick() { document.querySelectorAll('header, nav, [role="banner"], .header, .navbar, .sticky, .fixed-top, [style*="position: fixed"], [style*="position:sticky"]').forEach(function(el) { if (el.style.position === 'fixed' || el.style.position === 'sticky' || getComputedStyle(el).position === 'fixed' || getComputedStyle(el).position === 'sticky') { el.style.position = 'static'; el.style.top = 'auto'; el.style.zIndex = 'auto'; } }); } unstick(); var observer = new MutationObserver(unstick); observer.observe(document.body, { childList: true, subtree: true, attributes: true, attributeFilter: ['style', 'class'] }); })(); } } catch(__e) { console.warn('[Userscript:Kill Sticky Headers]', __e); } })(); })(); feat(agent): no-tools-called retry for local Ollama models (complements #210) by mike-011 · Pull Request #211 · VectifyAI/OpenKB · GitHub
Skip to content

feat(agent): no-tools-called retry for local Ollama models (complements #210) - #211

Open
mike-011 wants to merge 3 commits into
VectifyAI:mainfrom
mike-011:feat/ollama-no-tools-retry
Open

feat(agent): no-tools-called retry for local Ollama models (complements #210)#211
mike-011 wants to merge 3 commits into
VectifyAI:mainfrom
mike-011:feat/ollama-no-tools-retry

Conversation

@mike-011

@mike-011mike-011 commented Jul 29, 2026

Copy link
Copy Markdown

Summary

Complements PR #210 — adds a no-tools-called retry for local Ollama models (12-14B) that return a final answer without calling any tools.

Problem

After PR #210 fixed the ollama/ → ollama_chat/ rewrite and ModelBehaviorError retry, testing with local models revealed a second failure mode:

  • qwen2.5-coder:14b (local): add works (227s), but query returns "I cannot find relevant information" — the model didn't call read_file
  • gemma4:12b (local): query returns empty output ""

The KB is correctly built (wiki, concepts, entities all present), but the model doesn't initiate tool calls during query — it answers directly without reading wiki files.

This is a model capability limitation, not a transport bug. The adapter cannot force a model to call tools, but it can nudge it.

Fix

After Runner.run() completes successfully, the adapter checks result.new_items for any ToolCallItem. If the agent has tools but none were called, the adapter retries with an escalating nudge message:

Attempt 1 (gentle):

You answered without reading any wiki files. You have tools available. Call read_file with path 'summaries/index.md' first to see what documents exist, then answer the question.

Attempt 2 (firm):

You MUST use the read_file tool. Do NOT answer without reading files first. Start with read_file(path='summaries/index.md'). Available tools: read_file, get_page_content.

Attempt 3 (final):

IMPORTANT: Your previous answer was not grounded in the wiki. Call read_file now. The available tools are: read_file, get_page_content. Do not answer until you have called a tool.

If all retries are exhausted, the adapter returns the last (ungrounded) answer — no infinite loops.

Interaction with PR #210

This PR includes all changes from PR #210 (ollama_chat/ rewrite, ModelBehaviorError retry, parameterized timeouts, drop_params, api_base propagation) plus the no-tools-called retry. The two retry mechanisms share the same tool_call_retries budget.

Testing

  • 73 adapter unit tests (17 new for no-tools retry) — all passing
  • 1148 total tests — all passing (0 regressions)
python -m pytest tests/test_ollama_adapter.py -v # 73 passed
python -m pytest tests/ -q # 1148 passed

Files Changed

FileDescription
openkb/agent/ollama_adapter.pyAdded _has_tool_calls(), _build_nudge_message(), _append_nudge(), no-tools retry loop
tests/test_ollama_adapter.py17 new tests (73 total)
docs/ollama_no_tools_retry.mdNew — full documentation
(+ all files from PR #210)ollama_chat/ rewrite, compiler, cli, config, docs

Configuration

Same ollama: config block as PR #210:

ollama:
timeout: 300# per-request timeout (s), default 300tool_call_retries: 3# max retries (hallucination + no-tools), default 3

Limitations

  • The nudge cannot force a model to call tools — it can only encourage. Very small models (1B) may ignore it.
  • Models that produce empty output ("") may not benefit if the empty output is caused by a timeout or inference failure.
  • The retry budget is shared with ModelBehaviorError retries.

@mike-011
feat(agent): no-tools-called retry for local Ollama models
9f1e101
When a local Ollama model (12-14B) returns a final answer without calling
any tools, the adapter now retries with an escalating nudge message:
Attempt 1: 'You answered without reading wiki files. Call read_file
with path summaries/index.md first.'
Attempt 2: 'You MUST use the read_file tool. Do NOT answer without
reading files first.'
Attempt 3: 'IMPORTANT: Your previous answer was not grounded. Call
read_file now. Do not answer until you have called a tool.'
Detection: _has_tool_calls() checks result.new_items for ToolCallItem.
If none found and agent has tools, nudge and retry (up to tool_call_retries).
Bounded by tool_call_retries (default 3) — no infinite loops.
Shares retry budget with ModelBehaviorError retry.
Tests: 73 adapter tests (17 new), 1148 total, 0 failures.
Docs: docs/ollama_no_tools_retry.md
Complements PR VectifyAI#210 (ollama_chat/ rewrite + ModelBehaviorError retry).
@mike-011
fix(agent): sanitize ungrounded output after exhausted retries
06d875e
When no-tools-called retries are exhausted, the model output may contain
raw tool-call JSON (e.g. {name: read_file, ...}) or be empty.
Replace these with a clear user-facing message instead.
- Add _sanitize_ungrounded_output() to detect and replace raw JSON/empty
- Apply in both run_with_retry and arun_with_retry exhaustion paths
- Add 7 tests for sanitize behavior (80 adapter tests total)
- Document minimum model requirements based on live testing
Live test results:
qwen2.5-coder:14b — was returning raw JSON, now returns clear message
gemma4:12b — grounded answer (nudge retry works)
qwen3.5:397b-cloud — grounded answer (no retries needed)
@mike-011
mike-011force-pushed the feat/ollama-no-tools-retry branch from 3460dde to 06d875eCompareJuly 29, 2026 21:03
@mike-011
docs: minimum model requirements + updated test results
20a9349
- Add 'Minimum Model Requirements' section to ollama_tool_call_adapter.md
covering both add (low reqs) and query/chat (higher reqs) workloads
- Update tested models table with latest results including instability
notes for gemma4:12b and sanitize fallback for qwen2.5-coder:14b
- Update test counts (80 adapter, 1148 total)
- Cross-reference between the two docs
- No personal/test data in any commit
@mike-011

mike-011 commented Jul 29, 2026

Copy link
Copy Markdown
Author

Model Requirements & Sanitize Update

Minimum model requirements (documented in docs/ollama_tool_call_adapter.md)

OpenKB has two workloads with different LLM requirements:

add (ingestion) — direct LLM completion, no tool-calling needed. Models ≥12B work reliably.

query / chat — multi-step tool-calling loop (model must call read_file, process results, produce grounded answer). Requires strong instruction-following capability.

ModelSizeaddqueryNotes
qwen3.5 (cloud)397B✅ stable✅ groundedStable across all runs
gemma412B✅⚠️ unstableGrounded answer on some runs; "could not retrieve" on others
qwen2.5-coder14B✅❌Model ignores nudge retries; sanitize returns clear message

Sanitize fallback (new)

When no-tools retries are exhausted, raw JSON or empty output is replaced with:

"I could not find relevant information in the knowledge base. The available tools were not used successfully. Try rephrasing your question or adding more documents."

Recommendations

  • For query/chat: use general-purpose models (gemma4, qwen3), not code-focused (qwen2.5-coder)
  • For slow hardware: ollama.timeout: 600 or higher
  • 12B models may produce different results between runs — consider fixing temperature/seed

Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@mike-011