Module: PWN::MemoryIndex

Defined in:
lib/pwn/memory_index.rb

Overview

PWN::MemoryIndex is a lightweight local embedding index over PWN::Memory (~/.pwn/memory.json) so PromptBuilder can inject the N MOST-RELEVANT memories for the current request instead of the N newest. Embeddings come from the local Ollama instance (PWN::Env[:ollama][:embed_model], default 'nomic-embed-text') via its native /api/embed endpoint — everything stays on-box.

Index layout (~/.pwn/memory.idx):

{ "<key>": { "sha": "<sha16 of value>", "vec": [Float,…] }, … }

Rebuilds are incremental: only entries whose value-sha changed are (re)embedded, so a warm index costs one embed call (the query).

Constant Summary collapse

INDEX_FILE = 
File.join(Dir.home, '.pwn', 'memory.idx')
DEFAULT_EMBED_MODEL = 
'nomic-embed-text'

Class Method Summary collapse

Class Method Details

.authors ⇒ Object

Author(s)

0day Inc. support@0dayinc.com



221
222
223
 
# File 'lib/pwn/memory_index.rb', line 221

public_class_method def self.authors
  "AUTHOR(S):\n  0day Inc. <support@0dayinc.com>\n"
end

.available? ⇒ Boolean

Supported Method Parameters

bool = PWN::MemoryIndex.available?

True when a local Ollama base_uri is configured. All public methods degrade to substring recall when this is false.

Returns:

  • (Boolean) 


30
31
32
33
34
 
# File 'lib/pwn/memory_index.rb', line 30

public_class_method def self.available?
  !ollama_base.nil?
rescue StandardError
  false
end

.embed(opts = {}) ⇒ Object

Supported Method Parameters

vecs = PWN::MemoryIndex.embed(texts: ['a', 'b'])

POST /api/embed on the local Ollama; returns Array<Array> (one vector per input, nil on per-item failure). Batches of 32.



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
 
# File 'lib/pwn/memory_index.rb', line 128

public_class_method def self.embed(opts = {})
  texts = Array(opts[:texts]).map(&:to_s)
  base  = ollama_base
  return Array.new(texts.length) unless base && !texts.empty?

  model = (PWN::Env.dig(:ai, :ollama, :embed_model) if defined?(PWN::Env)) || DEFAULT_EMBED_MODEL
  browser = PWN::Plugins::TransparentBrowser.open(browser_type: :rest)
  rest    = browser[:browser]::Request
  headers = { content_type: 'application/json; charset=UTF-8' }
  token   = PWN::Env.dig(:ai, :ollama, :key) if defined?(PWN::Env)
  headers[:authorization] = "Bearer #{token}" if token

  out = []
  texts.each_slice(32) do |batch|
    resp = rest.execute(
      method: :post,
      url: "#{base}/ollama/api/embed",
      headers: headers,
      payload: { model: model, input: batch }.to_json,
      verify_ssl: false,
      timeout: 120
    )
    j = JSON.parse(resp, symbolize_names: true)
    out.concat(Array(j[:embeddings]))
  end
  out
rescue StandardError
  Array.new(texts.length)
end

.help ⇒ Object

Display Usage for this Module



227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
 
# File 'lib/pwn/memory_index.rb', line 227

public_class_method def self.help
  puts <<~USAGE
    USAGE:
      PWN::MemoryIndex.available?
      PWN::MemoryIndex.recall_semantic(query: 'nmap sweep', limit: 6)
      PWN::MemoryIndex.to_context(query: 'nmap sweep', limit: 6)  # PromptBuilder drop-in
      PWN::MemoryIndex.refresh                                    # incremental (re)embed
      PWN::MemoryIndex.embed(texts: ['a', 'b'])                   # raw vectors
      PWN::MemoryIndex.reset

      Config:
        PWN::Env[:ai][:ollama][:embed_model] = '<embed-model-tag>'  # any embedding model pulled locally

      #{self}.authors
  USAGE
end

.recall_semantic(opts = {}) ⇒ Object

Supported Method Parameters

hits = PWN::MemoryIndex.recall_semantic( query: 'required - user request / search text', limit: 'optional - top-K by cosine similarity (default 6)' )

Returns [{ key:, value:, category:, timestamp:, score: }, …] (newest-first Memory.recall shape + :score) or falls back to PWN::Memory.recall(query:) when embedding is unavailable.



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
 
# File 'lib/pwn/memory_index.rb', line 46

public_class_method def self.recall_semantic(opts = {})
  query = opts[:query].to_s
  limit = (opts[:limit] || 6).to_i
  mem   = PWN::Memory.load
  return [] if mem.empty? || query.strip.empty?

  qv = embed(texts: [query]).first
  return fallback(query: query, limit: limit) unless qv

  idx = refresh(mem: mem)
  now = Time.now.utc
  # M2 — Generative-Agents retrieval: score = α·sim + β·recency + γ·importance
  # (Park '23). pwn had only sim; recency and importance now let a
  # 3-day-old human :fact outrank a 1-hour-old heuristic :lesson.
  scored = mem.map do |k, v|
    vec = idx.dig(k, :vec)
    next unless vec

    sim = cosine(a: qv, b: vec)
    age_h = (now - Time.parse(v[:timestamp].to_s)) / 3_600.0
    rec = Math.exp(-age_h / (24.0 * 7.0))
    imp = (v[:importance] || 0.5).to_f
    score = (0.6 * sim) + (0.25 * rec) + (0.15 * imp)
    { key: k, value: v[:value], category: v[:category], timestamp: v[:timestamp], score: score, sim: sim.round(3), importance: imp }
  rescue StandardError
    nil
  end
  scored.compact.sort_by { |h| -h[:score] }.first(limit)
rescue StandardError
  fallback(query: query, limit: limit)
end

.refresh(opts = {}) ⇒ Object

Supported Method Parameters

idx = PWN::MemoryIndex.refresh(mem: 'optional - preloaded PWN::Memory hash')

Incrementally (re)embed changed entries and prune deleted keys. Returns the in-memory index Hash keyed by memory key (Symbol).



102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
 
# File 'lib/pwn/memory_index.rb', line 102

public_class_method def self.refresh(opts = {})
  mem = opts[:mem] || PWN::Memory.load
  idx = load_index
  idx.delete_if { |k, _| !mem.key?(k) }

  todo = mem.reject { |k, v| idx.dig(k, :sha) == sha(text: v[:value].to_s) }
  unless todo.empty?
    vecs = embed(texts: todo.values.map { |v| "#{v[:category]}: #{v[:value]}" })
    todo.keys.each_with_index do |k, i|
      next unless vecs[i]

      idx[k] = { sha: sha(text: mem[k][:value].to_s), vec: vecs[i] }
    end
    save_index(idx: idx)
  end
  idx
rescue StandardError
  idx || {}
end

.reset ⇒ Object

Supported Method Parameters

PWN::MemoryIndex.reset



161
162
163
164
 
# File 'lib/pwn/memory_index.rb', line 161

public_class_method def self.reset
  FileUtils.rm_f(INDEX_FILE)
  {}
end

.to_context(opts = {}) ⇒ Object

Supported Method Parameters

ctx = PWN::MemoryIndex.to_context(query:, limit: 6)

Drop-in replacement for PWN::Memory.to_context that ranks by relevance to query instead of insertion order. Emitted format is identical so PromptBuilder needs no special-casing.



85
86
87
88
89
90
91
92
93
94
 
# File 'lib/pwn/memory_index.rb', line 85

public_class_method def self.to_context(opts = {})
  hits = recall_semantic(query: opts[:query], limit: opts[:limit] || 6)
  return PWN::Memory.to_context(limit: opts[:limit] || 6) if hits.empty?

  ctx = "\n\nPERSISTENT MEMORY (relevance-ranked for this request - use PWN::Memory.remember to store new ones):\n"
  hits.each do |h|
    ctx += "- #{h[:key]} [#{h[:category]} @ #{h[:timestamp]}]: #{h[:value].to_s[0, 300]}\n"
  end
  ctx
end