Module: OllamaAgent::GemmaThoughtContentParser
- Defined in:
- lib/ollama_agent/gemma_thought_content_parser.rb
Overview
Incrementally strips Gemma-style reasoning channels from streamed message.content when Ollama does not populate message.thinking (common for Gemma 4 cloud vs Qwen/DeepSeek). merge_into_message_data! also runs ChatStreamThinkingFormat.normalize_message_thinking! so Hash/Array thinking payloads from the API are coerced to a String before display (README: Reasoning / thinking output).
Supported openings (earliest match wins):
-
<|channel>thought optionally followed by a single newline before reasoning text
-
<redacted_thinking>
Closures pair with the opening:
-
<channel|> (Gemma / Ollama template style)
-
</redacted_thinking>
rubocop:disable Metrics/ModuleLength – single streaming scanner; splitting would obscure state machine
Constant Summary collapse
- STATE_KEY =
"__ollama_agent_gemma_thought_parse"- OPEN_SPECS =
[ { key: "channel", open: "<|channel>thought", close: "<channel|>" }, { key: "redacted", open: "<redacted_thinking>", close: "</redacted_thinking>" } ].freeze
Class Method Summary collapse
- .attach_state!(carry, state) ⇒ Object
-
.extract_from_complete_content(content) ⇒ Array(String, String)
One-shot parse of a full assistant body (non-streaming chat / TUI).
-
.extract_state(carry) ⇒ Object
rubocop:enable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity.
- .initial_state ⇒ Object
-
.merge_into_message_data!(message) ⇒ Object
Mutates
messagebacking hash when the API omitsthinkingbut embeds Gemma channels incontent. -
.process_chunk(state, chunk) ⇒ Array(String, Hash, Array<String>)
rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity.
Class Method Details
.attach_state!(carry, state) ⇒ Object
102 103 104 105 106 |
# File 'lib/ollama_agent/gemma_thought_content_parser.rb', line 102 def attach_state!(carry, state) return unless carry.is_a?(Hash) carry[STATE_KEY] = copy_state(state) end |
.extract_from_complete_content(content) ⇒ Array(String, String)
One-shot parse of a full assistant body (non-streaming chat / TUI).
110 111 112 113 114 115 116 117 118 119 |
# File 'lib/ollama_agent/gemma_thought_content_parser.rb', line 110 def extract_from_complete_content(content) c = content.to_s return [nil, c] if c.empty? new_content, _state, deltas = process_chunk(initial_state, c) thinking = deltas.join return [nil, c] if thinking.strip.empty? [thinking, new_content] end |
.extract_state(carry) ⇒ Object
rubocop:enable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity
93 94 95 96 97 98 99 100 |
# File 'lib/ollama_agent/gemma_thought_content_parser.rb', line 93 def extract_state(carry) return initial_state unless carry.is_a?(Hash) raw = carry[STATE_KEY] return initial_state if raw.nil? || !raw.is_a?(Hash) copy_state(raw) end |
.initial_state ⇒ Object
29 30 31 32 33 34 35 |
# File 'lib/ollama_agent/gemma_thought_content_parser.rb', line 29 def initial_state { "phase" => "content", "pend" => +"", "open_key" => nil } end |
.merge_into_message_data!(message) ⇒ Object
Mutates message backing hash when the API omits thinking but embeds Gemma channels in content.
122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 |
# File 'lib/ollama_agent/gemma_thought_content_parser.rb', line 122 def () data = () return unless data ChatStreamThinkingFormat.(data) return if native_thinking_present?(data) raw_content = (data["content"] || data[:content]).to_s return if raw_content.empty? thinking_text, visible = extract_from_complete_content(raw_content) return if thinking_text.nil? || thinking_text.strip.empty? data["thinking"] = thinking_text data["content"] = visible end |
.process_chunk(state, chunk) ⇒ Array(String, Hash, Array<String>)
rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity
39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 |
# File 'lib/ollama_agent/gemma_thought_content_parser.rb', line 39 def process_chunk(state, chunk) state = copy_state(state) return [nil, state, []] if chunk.nil? deltas = [] content_out = +"" work = state["pend"] + chunk state["pend"] = +"" until work.empty? if state["phase"] == "content" earliest = earliest_open(work) if earliest.nil? hold = longest_open_prefix_suffix(work) emit_len = work.length - hold.length content_out << work[0, emit_len] if emit_len.positive? state["pend"] = hold break end idx, spec = earliest olen = spec[:open].length content_out << work[0, idx] if idx.positive? work = work[(idx + olen)..] || +"" work = work[1..] if work.start_with?("\n") state["phase"] = "thought" state["open_key"] = spec[:key] next end spec = OPEN_SPECS.find { |s| s[:key] == state["open_key"] } idx = work.index(spec[:close]) if idx.nil? hold = longest_close_prefix_suffix(work, spec[:close]) emit_len = work.length - hold.length part = emit_len.positive? ? work[0, emit_len] : +"" deltas << part if part != "" state["pend"] = hold break end clen = spec[:close].length part = work[0, idx] deltas << part if part != "" work = work[(idx + clen)..] || +"" work = work[1..] if work.start_with?("\n") state["phase"] = "content" state["open_key"] = nil end [content_out, state, deltas] end |