Module: OllamaChat::CollectionSearch

Includes:
Kramdown::ANSI::Width
Included in:
Chat
Defined in:
lib/ollama_chat/collection_search.rb

Overview

Shared collection search primitives for document retrieval and LLM reranking.

Extracted from Tools::SearchKnowledge so that the /collection query command and the trigger mechanism can reuse the same search + rerank pipeline without duplicating logic.

Instance Method Summary collapse

Instance Method Details

#find_document_records(query, tags: nil, text_size: nil, text_count: nil, min_similarity: nil) ⇒ Array<Documentrix::Utils::TagResult>

Searches the active document collection for records matching the given query.

Parameters:

  • query (String) —

    the search query

  • tags (Array<String>, nil) (defaults to: nil) —

    tag filter

  • text_size (Integer, nil) (defaults to: nil) —

    max snippet size

  • text_count (Integer, nil) (defaults to: nil) —

    max number of snippets

  • min_similarity (Float, nil) (defaults to: nil) —

    minimum similarity threshold

Returns:

  • (Array<Documentrix::Utils::TagResult>) —

    matching records



20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
 
# File 'lib/ollama_chat/collection_search.rb', line 20

def find_document_records(
  query,
  tags: nil, text_size: nil, text_count: nil, min_similarity: nil
)
  tags = Documentrix::Utils::Tags.new(
    tags, valid_tag: /\A#*([-\w.\]\[]+)/
  )
  effective_query  = query.first(
    config.embedding.model.context_length
  )
  embedding_prompt = config.embedding.model.prompt?
  log(
    :info, 'Embedding search request',
    data: {
      url:        embedding_ollama.base_url,
      model:      config.embedding.model.to_h,
      collection: collection,
      query:      effective_query,
      prompt:     embedding_prompt,
    }
  )
  records = @documents.find_where(
    effective_query,
    tags:,
    prompt:         embedding_prompt,
    text_size:,
    text_count:,
    min_similarity:
  )
  log(
    :info, 'Embedding search response',
    data: {
      url:          embedding_ollama.base_url,
      collection:   collection,
      query:        effective_query,
      result_count: records.size,
      top:          records.first&.similarity&.round(4),
    }
  )
  records
end

#rerank_records(query, records, prompt_name: nil) ⇒ Array<Documentrix::Utils::TagResult>

Reranks a set of records using the LLM with the rerank prompt.

Parameters:

  • query (String) —

    the original search query

  • records (Array<Documentrix::Utils::TagResult>) —

    candidate records

  • prompt_name (String, nil) (defaults to: nil) —

    the prompt template to use for reranking; defaults to 'rerank'

Returns:

  • (Array<Documentrix::Utils::TagResult>) —

    filtered records



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_chat/collection_search.rb', line 69

def rerank_records(query, records, prompt_name: nil)
  candidates = records.each_with_index.map { |r, i|
    "[#{i}] #{r.tags_set.to_s(link: false)} #{truncate(r.text.squeeze, length: 300)}"
  }.join("\n")

  prompt_name ||= 'default'
  rerank_prompt = prompt(prompt_name, context: 'rerank') or
    raise 'missing prompt %s' % prompt_name.inspect
  rerank_prompt = rerank_prompt.to_s
    .named_placeholders_interpolate({ query:, candidates: })

  begin
    if response = generate(prompt: rerank_prompt).full?
      indices = response.scan(/\d+/).map(&:to_i)
        .select { |i| (0...records.size).include?(i) }
      records = records.values_at(*indices) if indices.any?
    end
  rescue => e
    log(:error, e, data: { context: 'rerank' })
  end
  records
end

#trigger_inject(content, group_uuid:) ⇒ self?

Injects relevant snippets from the session's trigger collection into the conversation as a tool message.

Reads the first entry of the session's trigger hash, performs a vector search in the configured collection, reranks the candidates via the LLM, and appends a trigger_inject tool message containing the surviving snippets (text, similarity, tags). The message is grouped with the current user message via group_uuid.

Parameters:

  • content (String) —

    the parsed user message text used as the search query

  • group_uuid (String, nil) —

    the group UUID for message grouping (shared with the user message and runtime-info message)

Returns:

  • (self, nil) —

    self if snippets were injected, nil if no trigger is configured, the trigger or embedding is disabled, the collection or prompt is missing, or no snippets survived reranking



109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
 
# File 'lib/ollama_chat/collection_search.rb', line 109

def trigger_inject(content, group_uuid:)
  embedding.on? or return
  trigger = session.trigger&.first or return
  collection, config = trigger
  unless config['enabled']
    log(:info, 'Trigger for collection %s disabled.' % collection)
    return
  end
  unless col = database_collection?(collection)
    log(:error, 'Unknown collection named %s' % collection)
    return
  end
  prompt_name = config['prompt_name']
  unless rerank_prompt = prompt(prompt_name, context: 'rerank').full?(:to_s)
    log(
      :error, 'Unknown rerank prompt named %s' % prompt_name,
      data: { config: }
    )
    return
  end
  records = []
  switch_collection(collection) do
    records = find_document_records(content, text_count: config['text_count'])
    records.empty? and return
    pre_rerank = records.size
    records    = rerank_records(content, records, prompt_name:)
    log(
      :info,
      'Trigger: %d/%d passed rerank for %s' % [ records.size, pre_rerank, collection ],
      data: {
        collection:,
        prompt_name:,
        records: records.map { |r|
          {
            text:       r.text,
            similarity: r.similarity&.round(4),
            tags:       r.tags_set.to_s(link: false)
          }
        }
      })
    records.empty? and return
  end
  snippets = records.map { |record|
    {
      text:       record.text,
      similarity: record.similarity.to_f,
      tags:       record.tags_set.to_s(link: false),
    }
  }
  message_content = {
    prompt: prompt('snippets_trigger').to_s,
    collection: {
      name:        collection,
      description: col.description&.to_s,
    },
    snippets:
  }.to_json
  tool_name       = 'trigger_inject'
  messages << OllamaChat::Message.new(
    role:        'user',
    tool_name:   ,
    sender_name: tool_name,
    content:     message_content,
    group_uuid:
  )
  self
end